llama/parsers/renderers: nemotron 3 nano (#13489)

---------
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

llama/llama.cpp/src/llama-arch.cpp

@@ -75,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
+    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -1765,6 +1766,39 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NEMOTRON_H_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // MoE FFN (for MoE layers)
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,"blk.%d.exp_probs_b" },
+            // MoE shared expert layer
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_EXAONE,
         {
@@ -2838,6 +2872,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_QWEN3NEXT:
             return true;
         default:

llama/llama.cpp/src/llama-arch.h

@@ -79,6 +79,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
+    LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,

llama/llama.cpp/src/llama-graph.cpp

@@ -1089,6 +1089,16 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_moe_relu", il);
             } break;
+        case LLM_FFN_RELU_SQR:
+            if (gate_exps) {
+                // TODO: add support for gated squared relu
+                GGML_ABORT("fatal error: gated squared relu not implemented");
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_moe_relu_sqr", il);
+            }
+            break;
         default:
             GGML_ABORT("fatal error");
     }

llama/llama.cpp/src/llama-model.cpp

@@ -120,6 +120,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
+        case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
@@ -1788,6 +1790,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -1803,7 +1806,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+
                 switch (hparams.n_layer) {
+                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -5175,6 +5185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_NEMOTRON_H:
+            case LLM_ARCH_NEMOTRON_H_MOE:
                 {
                     // mamba2 Mixer SSM params
                     // NOTE: int64_t for tensor dimensions
@@ -5185,6 +5196,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5234,6 +5248,19 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
+                        }  else {
+                            if (n_expert != 0) {
+                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                                // MoE branch
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                                // Shared expert branch
+                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
                             } else {
                                 // mlp layers
                                 layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
@@ -5242,6 +5269,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                                 layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
                             }
                         }
+                    }
                 } break;
             case LLM_ARCH_EXAONE:
                 {
@@ -6870,7 +6898,8 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
         arch == LLM_ARCH_QWEN3NEXT ||
-        arch == LLM_ARCH_NEMOTRON_H) {
+        arch == LLM_ARCH_NEMOTRON_H ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -6926,7 +6955,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
         arch == LLM_ARCH_GRANITE_MOE ||
-        arch == LLM_ARCH_GRANITE_HYBRID) {
+        arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7107,7 +7137,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](int32_t) { return true; };
                         filter_recr = [&](int32_t) { return true; };
-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
+                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
                         filter_attn = [&](int32_t il) {
                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
                         };
@@ -7478,6 +7508,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
             } break;
@@ -7765,6 +7796,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARWKV7:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values

llama/llama.cpp/src/llama-model.h

@@ -114,6 +114,7 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_31B_A3_5B,
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air

llama/llama.cpp/src/models/nemotron-h.cpp

@@ -107,12 +107,41 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
 }
 
 ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
+    if (model.layers[il].ffn_gate_inp == nullptr) {
         cur = build_ffn(cur,
                 model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                 NULL,                      NULL,                        NULL,
                 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-            NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
         cb(cur, "ffn_out", il);
+    } else {
+        ggml_tensor * ffn_inp = cur;
+        ggml_tensor * moe_out =
+            build_moe_ffn(ffn_inp,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    nullptr, // no gate
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up_shexp,  NULL, NULL,
+                    NULL /* no gate */           ,  NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "ffn_shexp", il);
+
+        cur = ggml_add(ctx0, moe_out, ffn_shexp);
+        cb(cur, "ffn_out", il);
+    }
 
     cur = build_cvec(cur, il);
     cb(cur, "l_out", il);

model/parsers/nemotron3nano.go

+package parsers
+
+import (
+	"regexp"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+)
+
+type Nemotron3NanoParserState int
+
+const (
+	Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
+	Nemotron3NanoSkipWhitespaceAfterThinking
+	Nemotron3NanoCollectingContent
+	Nemotron3NanoCollectingToolCalls
+)
+
+const (
+	nemotronThinkClose    = "</think>"
+	nemotronToolCallOpen  = "<tool_call>"
+	nemotronToolCallClose = "</tool_call>"
+)
+
+type Nemotron3NanoParser struct {
+	state       Nemotron3NanoParserState
+	buffer      strings.Builder
+	tools       []api.Tool
+	HasThinking bool
+}
+
+func (p *Nemotron3NanoParser) HasToolSupport() bool     { return true }
+func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return p.HasThinking }
+
+func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.tools = tools
+
+	// Check both model capability AND request preference
+	thinkingEnabled := thinkValue != nil && thinkValue.Bool()
+
+	prefill := lastMessage != nil && lastMessage.Role == "assistant"
+
+	if !thinkingEnabled {
+		p.state = Nemotron3NanoCollectingContent
+		return tools
+	}
+
+	if prefill && lastMessage.Content != "" {
+		p.state = Nemotron3NanoCollectingContent
+		return tools
+	}
+
+	p.state = Nemotron3NanoCollectingThinking
+	return tools
+}
+
+type nemotronEvent interface {
+	isNemotronEvent()
+}
+
+type nemotronEventThinkingContent struct {
+	content string
+}
+
+type nemotronEventContent struct {
+	content string
+}
+
+type nemotronEventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (nemotronEventThinkingContent) isNemotronEvent() {}
+func (nemotronEventContent) isNemotronEvent()         {}
+func (nemotronEventToolCall) isNemotronEvent()        {}
+
+func (p *Nemotron3NanoParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+	for _, event := range events {
+		switch event := event.(type) {
+		case nemotronEventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case nemotronEventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case nemotronEventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
+	var all []nemotronEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []nemotronEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+// emitWithPartialCheck extracts unambiguous content before a potential partial tag
+func (p *Nemotron3NanoParser) emitWithPartialCheck(bufStr, tag string) (unambiguous, ambiguous string) {
+	if overlapLen := overlap(bufStr, tag); overlapLen > 0 {
+		beforePartialTag := bufStr[:len(bufStr)-overlapLen]
+		trailingLen := trailingWhitespaceLen(beforePartialTag)
+		return bufStr[:len(beforePartialTag)-trailingLen], bufStr[len(beforePartialTag)-trailingLen:]
+	}
+	wsLen := trailingWhitespaceLen(bufStr)
+	return bufStr[:len(bufStr)-wsLen], bufStr[len(bufStr)-wsLen:]
+}
+
+func (p *Nemotron3NanoParser) eat() ([]nemotronEvent, bool) {
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return nil, false
+	}
+
+	switch p.state {
+	case Nemotron3NanoCollectingThinking:
+		if strings.Contains(bufStr, nemotronThinkClose) {
+			split := strings.SplitN(bufStr, nemotronThinkClose, 2)
+			thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			remainder := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.WriteString(remainder)
+			// Transition to whitespace-skipping state if buffer is empty,
+			// otherwise go directly to content collection
+			if remainder == "" {
+				p.state = Nemotron3NanoSkipWhitespaceAfterThinking
+			} else {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			if thinking != "" {
+				return []nemotronEvent{nemotronEventThinkingContent{content: thinking}}, true
+			}
+			return nil, true
+		}
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronThinkClose)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
+		}
+		return nil, false
+
+	// We only want to skip whitespace between thinking and content
+	case Nemotron3NanoSkipWhitespaceAfterThinking:
+		bufStr = strings.TrimLeftFunc(bufStr, unicode.IsSpace)
+		p.buffer.Reset()
+		p.buffer.WriteString(bufStr)
+		if bufStr == "" {
+			return nil, false
+		}
+		p.state = Nemotron3NanoCollectingContent
+		return nil, true
+
+	case Nemotron3NanoCollectingContent:
+		if strings.Contains(bufStr, nemotronToolCallOpen) {
+			split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
+			content := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(split[1])
+			p.state = Nemotron3NanoCollectingToolCalls
+			if content != "" {
+				return []nemotronEvent{nemotronEventContent{content: content}}, true
+			}
+			return nil, true
+		}
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventContent{content: unambig}}, false
+		}
+		return nil, false
+
+	case Nemotron3NanoCollectingToolCalls:
+		if strings.Contains(bufStr, nemotronToolCallClose) {
+			split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
+			remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+
+			var events []nemotronEvent
+			if tc, err := p.parseToolCall(split[0]); err == nil {
+				events = append(events, nemotronEventToolCall{toolCall: tc})
+			}
+
+			if !strings.Contains(remaining, nemotronToolCallOpen) {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			return events, true
+		}
+		return nil, false
+	}
+
+	return nil, false
+}
+
+var (
+	nemotronFunctionRegex  = regexp.MustCompile(`<function=([^>]+)>`)
+	nemotronParameterRegex = regexp.MustCompile(`<parameter=([^>]+)>\n?([\s\S]*?)\n?</parameter>`)
+)
+
+func (p *Nemotron3NanoParser) parseToolCall(content string) (api.ToolCall, error) {
+	toolCall := api.ToolCall{}
+
+	// Extract function name
+	fnMatch := nemotronFunctionRegex.FindStringSubmatch(content)
+	if len(fnMatch) < 2 {
+		return toolCall, nil
+	}
+	toolCall.Function.Name = fnMatch[1]
+
+	// Extract parameters
+	toolCall.Function.Arguments = make(api.ToolCallFunctionArguments)
+	paramMatches := nemotronParameterRegex.FindAllStringSubmatch(content, -1)
+	for _, match := range paramMatches {
+		if len(match) >= 3 {
+			paramName := match[1]
+			paramValue := strings.TrimSpace(match[2])
+
+			// Try to parse as typed value based on tool definition
+			toolCall.Function.Arguments[paramName] = p.parseParamValue(paramName, paramValue)
+		}
+	}
+
+	return toolCall, nil
+}
+
+func (p *Nemotron3NanoParser) parseParamValue(paramName string, raw string) any {
+	// Find the matching tool to get parameter type
+	var paramType api.PropertyType
+	for _, tool := range p.tools {
+		if prop, ok := tool.Function.Parameters.Properties[paramName]; ok {
+			paramType = prop.Type
+			break
+		}
+	}
+
+	return parseValue(raw, paramType)
+}

model/parsers/nemotron3nano_test.go

+package parsers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestNemotron3NanoParser(t *testing.T) {
+	tests := []struct {
+		name             string
+		input            string
+		thinkValue       *api.ThinkValue
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+	}{
+		{
+			name:            "simple content - no thinking",
+			input:           "Hello, how can I help you?",
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:            "simple content - thinking disabled",
+			input:           "Hello, how can I help you?",
+			thinkValue:      &api.ThinkValue{Value: false},
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:             "thinking then content",
+			input:            "Let me think about this...</think>\nHere is my answer.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think about this...",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:             "thinking with newlines",
+			input:            "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude</think>\nThe answer is 42.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude",
+			expectedContent:  "The answer is 42.",
+		},
+		{
+			name:       "simple tool call",
+			input:      "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:            "content then tool call",
+			input:           "Let me check the weather.\n<tool_call>\n<function=get_weather>\n<parameter=city>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiple parameters",
+			input:      "<tool_call>\n<function=book_flight>\n<parameter=from>\nSFO\n</parameter>\n<parameter=to>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls",
+			input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nSan Francisco\n</parameter>\n</function>\n</tool_call>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nNew York\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "San Francisco"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "New York"},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking then tool call",
+			input:            "I should check the weather...</think>\n<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "I should check the weather...",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking content then tool call",
+			input:            "Let me think...</think>\nI'll check for you.\n<tool_call>\n<function=search>\n<parameter=query>\ntest\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think...",
+			expectedContent:  "I'll check for you.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": "test"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiline parameter value",
+			input:      "<tool_call>\n<function=create_note>\n<parameter=content>\nLine 1\nLine 2\nLine 3\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
+					},
+				},
+			},
+		},
+		{
+			name:             "empty thinking block - immediate close",
+			input:            "</think>\nHere is my answer.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:            "thinking disabled but model outputs think close anyway",
+			input:           "</think>\nSome content after spurious tag.",
+			thinkValue:      &api.ThinkValue{Value: false},
+			expectedContent: "</think>\nSome content after spurious tag.",
+		},
+		{
+			name:          "tool call with no function name - returns empty tool call",
+			input:         "<tool_call>\n<function=>\n</function>\n</tool_call>",
+			thinkValue:    nil,
+			expectedCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "", Arguments: nil}}},
+		},
+		{
+			name:            "content with newlines preserved",
+			input:           "Line 1\n\nLine 2\n\n\nLine 3",
+			thinkValue:      nil,
+			expectedContent: "Line 1\n\nLine 2\n\n\nLine 3",
+		},
+		{
+			name:             "thinking with only whitespace after close tag",
+			input:            "My thoughts...</think>   \n\t\n   Content here.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "My thoughts...",
+			expectedContent:  "Content here.",
+		},
+		{
+			name:            "unicode content",
+			input:           "Hello 世界! 🌍 Ñoño",
+			thinkValue:      nil,
+			expectedContent: "Hello 世界! 🌍 Ñoño",
+		},
+		{
+			name:       "tool call with numeric parameter",
+			input:      "<tool_call>\n<function=set_temp>\n<parameter=value>\n42\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "set_temp",
+						Arguments: map[string]any{"value": "42"},
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Nemotron3NanoParser{HasThinking: tt.thinkValue != nil && tt.thinkValue.Bool()}
+			p.Init(nil, nil, tt.thinkValue)
+
+			content, thinking, calls, err := p.Add(tt.input, false)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			// Drain remaining content
+			finalContent, finalThinking, finalCalls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			content += finalContent
+			thinking += finalThinking
+			calls = append(calls, finalCalls...)
+
+			if diff := cmp.Diff(content, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(thinking, tt.expectedThinking); diff != "" {
+				t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(calls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestNemotron3NanoParser_Streaming(t *testing.T) {
+	tests := []struct {
+		name             string
+		chunks           []string
+		thinkValue       *api.ThinkValue
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+	}{
+		{
+			name:            "streaming content character by character",
+			chunks:          []string{"H", "e", "l", "l", "o", ",", " ", "w", "o", "r", "l", "d", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello, world!",
+		},
+		{
+			name:            "streaming content small tokens",
+			chunks:          []string{"Hel", "lo", ", ", "how ", "can", " I", " help", " you", " today", "?"},
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you today?",
+		},
+		{
+			name:             "streaming thinking then content - granular",
+			chunks:           []string{"Let", " me", " th", "ink", " about", " this", "...", "<", "/", "think", ">", "\n", "Here", " is", " my", " answer", "."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think about this...",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:             "streaming thinking with newlines - granular",
+			chunks:           []string{"Step", " 1", ":", " Ana", "lyze\n", "Step", " 2", ":", " Pro", "cess", "</", "thi", "nk>", "\n", "The", " ans", "wer."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Step 1: Analyze\nStep 2: Process",
+			expectedContent:  "The answer.",
+		},
+		{
+			name:       "streaming tool call - highly granular",
+			chunks:     []string{"<", "tool", "_", "call", ">", "\n", "<", "func", "tion", "=", "get", "_", "weather", ">", "\n", "<", "param", "eter", "=", "city", ">", "\n", "Par", "is", "\n", "</", "param", "eter", ">", "\n", "</", "func", "tion", ">", "\n", "</", "tool", "_", "call", ">"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:            "streaming content then tool call - granular",
+			chunks:          []string{"Let", " me", " check", " the", " weather", ".", "\n<", "tool_call", ">", "\n", "<function=", "get_weather", ">", "\n", "<parameter=", "city", ">", "\n", "NYC", "\n", "</parameter>", "\n", "</function>", "\n", "</tool_call>"},
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:   "tool call tag split character by character",
+			chunks: []string{"<", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">", "\n", "<", "f", "u", "n", "c", "t", "i", "o", "n", "=", "t", "e", "s", "t", ">", "\n", "<", "/", "f", "u", "n", "c", "t", "i", "o", "n", ">", "\n", "<", "/", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">"},
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking close tag split character by character",
+			chunks:           []string{"I", "'", "m", " ", "t", "h", "i", "n", "k", "i", "n", "g", ".", ".", ".", "<", "/", "t", "h", "i", "n", "k", ">", "\n", "D", "o", "n", "e", "!"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "I'm thinking...",
+			expectedContent:  "Done!",
+		},
+		{
+			name:             "multiple whitespace after think tag - separate chunks",
+			chunks:           []string{"Thinking...", "</think>", "\n", "\n", " ", "Content here."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Thinking...",
+			expectedContent:  "Content here.",
+		},
+		{
+			name:       "tool call with multiple parameters - streaming",
+			chunks:     []string{"<tool_", "call>\n", "<function", "=book_", "flight>", "\n<para", "meter=", "from>\n", "SFO\n", "</param", "eter>", "\n<param", "eter=to", ">\nNYC", "\n</para", "meter>", "\n</func", "tion>\n", "</tool_", "call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking then content then tool call - streaming",
+			chunks:           []string{"Ana", "lyzing", " your", " request", "...", "</", "think", ">\n", "I'll", " check", " that", " for", " you", ".", "\n", "<tool", "_call", ">\n", "<function", "=search", ">\n", "<parameter", "=query", ">\n", "test", " query", "\n</", "parameter", ">\n", "</function", ">\n", "</tool", "_call", ">"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Analyzing your request...",
+			expectedContent:  "I'll check that for you.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": "test query"},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls - streaming",
+			chunks: []string{
+				"<tool_call>", "\n", "<function=", "get_weather>", "\n",
+				"<parameter=", "city>\n", "San Fran", "cisco\n", "</parameter>", "\n",
+				"</function>", "\n", "</tool_call>", "\n",
+				"<tool_", "call>\n", "<function", "=get_weather", ">\n",
+				"<param", "eter=city", ">\nNew", " York\n", "</parameter>\n",
+				"</function>\n", "</tool_call>",
+			},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "San Francisco"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "New York"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiline parameter - streaming",
+			chunks:     []string{"<tool_call>\n", "<function=", "create_note>\n", "<parameter=", "content>\n", "Line 1", "\nLine", " 2\n", "Line 3", "\n</parameter>\n", "</function>\n", "</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
+					},
+				},
+			},
+		},
+		{
+			name:             "empty thinking block",
+			chunks:           []string{"</think>", "\n", "Just content."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "",
+			expectedContent:  "Just content.",
+		},
+		{
+			name:            "empty input chunks interspersed",
+			chunks:          []string{"Hello", "", " ", "", "world", "", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello world!",
+		},
+		{
+			name:             "tool call immediately after think close - no content",
+			chunks:           []string{"Analyzing...", "</think>", "\n", "<tool_call>", "\n<function=test>\n</function>\n", "</tool_call>"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Analyzing...",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with empty parameter value",
+			chunks:     []string{"<tool_call>\n<function=test>\n<parameter=name>\n", "\n</parameter>\n</function>\n</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{"name": ""},
+					},
+				},
+			},
+		},
+		{
+			name:            "partial tool call tag at end - buffered",
+			chunks:          []string{"Here's some content", "<tool"},
+			thinkValue:      nil,
+			expectedContent: "Here's some content",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Nemotron3NanoParser{HasThinking: tt.thinkValue != nil && tt.thinkValue.Bool()}
+			p.Init(nil, nil, tt.thinkValue)
+
+			var allContent string
+			var allThinking string
+			var allCalls []api.ToolCall
+
+			for _, chunk := range tt.chunks {
+				content, thinking, calls, err := p.Add(chunk, false)
+				if err != nil {
+					t.Fatalf("unexpected error: %v", err)
+				}
+				allContent += content
+				allThinking += thinking
+				allCalls = append(allCalls, calls...)
+			}
+
+			// Drain
+			content, thinking, calls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			allContent += content
+			allThinking += thinking
+			allCalls = append(allCalls, calls...)
+
+			if diff := cmp.Diff(allContent, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(allThinking, tt.expectedThinking); diff != "" {
+				t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(allCalls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestNemotron3NanoParser_HasToolSupport(t *testing.T) {
+	p := &Nemotron3NanoParser{}
+	if !p.HasToolSupport() {
+		t.Error("expected HasToolSupport to return true")
+	}
+}
+
+func TestNemotron3NanoParser_HasThinkingSupport(t *testing.T) {
+	t.Run("with thinking enabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: true}
+		if !p.HasThinkingSupport() {
+			t.Error("expected HasThinkingSupport to return true")
+		}
+	})
+
+	t.Run("with thinking disabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: false}
+		if p.HasThinkingSupport() {
+			t.Error("expected HasThinkingSupport to return false")
+		}
+	})
+}
+
+func TestNemotron3NanoParser_Init(t *testing.T) {
+	t.Run("starts in thinking state when enabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: true}
+		p.Init(nil, nil, &api.ThinkValue{Value: true})
+		if p.state != Nemotron3NanoCollectingThinking {
+			t.Errorf("expected state Nemotron3NanoCollectingThinking, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state when thinking disabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: true}
+		p.Init(nil, nil, &api.ThinkValue{Value: false})
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state when nil thinkValue", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: true}
+		p.Init(nil, nil, nil)
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state with assistant prefill", func(t *testing.T) {
+		p := &Nemotron3NanoParser{HasThinking: true}
+		prefill := &api.Message{Role: "assistant", Content: "Starting..."}
+		p.Init(nil, prefill, &api.ThinkValue{Value: true})
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+}
+
+func TestNemotron3NanoParser_WithTools(t *testing.T) {
+	tools := []api.Tool{
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name: "get_weather",
+				Parameters: api.ToolFunctionParameters{
+					Type: "object",
+					Properties: map[string]api.ToolProperty{
+						"city": {Type: api.PropertyType{"string"}},
+					},
+				},
+			},
+		},
+	}
+
+	p := &Nemotron3NanoParser{}
+	returnedTools := p.Init(tools, nil, nil)
+
+	if diff := cmp.Diff(returnedTools, tools); diff != "" {
+		t.Errorf("tools mismatch (-got +want):\n%s", diff)
+	}
+
+	// Parse a tool call
+	input := "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>"
+	_, _, calls, err := p.Add(input, true)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expectedCalls := []api.ToolCall{
+		{
+			Function: api.ToolCallFunction{
+				Name:      "get_weather",
+				Arguments: map[string]any{"city": "Paris"},
+			},
+		},
+	}
+
+	if diff := cmp.Diff(calls, expectedCalls); diff != "" {
+		t.Errorf("calls mismatch (-got +want):\n%s", diff)
+	}
+}

model/parsers/parsers.go

@@ -62,6 +62,10 @@ func ParserForName(name string) Parser {
 		return &Olmo3Parser{}
 	case "olmo3-think":
 		return &Olmo3ThinkParser{}
+	case "nemotron-3-nano":
+		return &Nemotron3NanoParser{HasThinking: false}
+	case "nemotron-3-nano-thinking":
+		return &Nemotron3NanoParser{HasThinking: true}
 	default:
 		return nil
 	}

model/renderers/nemotron3nano.go

+package renderers
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type Nemotron3NanoRenderer struct {
+	IsThinking bool
+}
+
+func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	// thinking is enabled: model must support it AND user must request it
+	enableThinking := r.IsThinking && (thinkValue != nil && thinkValue.Bool())
+
+	// Extract system message if present
+	var systemMessage string
+	var loopMessages []api.Message
+	if len(messages) > 0 && messages[0].Role == "system" {
+		systemMessage = messages[0].Content
+		loopMessages = messages[1:]
+	} else {
+		loopMessages = messages
+	}
+
+	// Find last user message index for thinking truncation
+	lastUserIdx := -1
+	for i, msg := range loopMessages {
+		if msg.Role == "user" {
+			lastUserIdx = i
+		}
+	}
+
+	sb.WriteString("<|im_start|>system\n")
+	if systemMessage != "" {
+		sb.WriteString(systemMessage)
+	}
+
+	if len(tools) > 0 {
+		if systemMessage != "" {
+			sb.WriteString("\n\n")
+		}
+		sb.WriteString(r.renderTools(tools))
+	}
+	sb.WriteString("<|im_end|>\n")
+
+	for i, message := range loopMessages {
+		switch message.Role {
+		case "assistant":
+			// Build content with thinking tags
+			content := r.buildContent(message)
+			shouldTruncate := i < lastUserIdx
+
+			if len(message.ToolCalls) > 0 {
+				sb.WriteString("<|im_start|>assistant\n")
+				sb.WriteString(r.formatContent(content, shouldTruncate, true))
+				r.writeToolCalls(&sb, message.ToolCalls)
+				sb.WriteString("<|im_end|>\n")
+			} else {
+				formatted := r.formatContent(content, shouldTruncate, false)
+				sb.WriteString("<|im_start|>assistant\n" + formatted + "<|im_end|>\n")
+			}
+
+		case "user", "system":
+			sb.WriteString("<|im_start|>" + message.Role + "\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+
+		case "tool":
+			// Check if previous message was also a tool message
+			prevWasTool := i > 0 && loopMessages[i-1].Role == "tool"
+			nextIsTool := i+1 < len(loopMessages) && loopMessages[i+1].Role == "tool"
+
+			if !prevWasTool {
+				sb.WriteString("<|im_start|>user\n")
+			}
+			sb.WriteString("<tool_response>\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("\n</tool_response>\n")
+
+			if !nextIsTool {
+				sb.WriteString("<|im_end|>\n")
+			}
+
+		default:
+			sb.WriteString("<|im_start|>" + message.Role + "\n" + message.Content + "<|im_end|>\n")
+		}
+	}
+
+	// Add generation prompt
+	if enableThinking {
+		sb.WriteString("<|im_start|>assistant\n<think>\n")
+	} else {
+		sb.WriteString("<|im_start|>assistant\n<think></think>")
+	}
+
+	return sb.String(), nil
+}
+
+func (r *Nemotron3NanoRenderer) renderTools(tools []api.Tool) string {
+	var sb strings.Builder
+	sb.WriteString("# Tools\n\nYou have access to the following functions:\n\n<tools>")
+
+	for _, tool := range tools {
+		fn := tool.Function
+		sb.WriteString("\n<function>\n<name>" + fn.Name + "</name>")
+
+		if fn.Description != "" {
+			sb.WriteString("\n<description>" + strings.TrimSpace(fn.Description) + "</description>")
+		}
+
+		sb.WriteString("\n<parameters>")
+		if fn.Parameters.Properties != nil {
+			for paramName, paramFields := range fn.Parameters.Properties {
+				sb.WriteString("\n<parameter>")
+				sb.WriteString("\n<name>" + paramName + "</name>")
+
+				if len(paramFields.Type) > 0 {
+					sb.WriteString("\n<type>" + strings.Join(paramFields.Type, ", ") + "</type>")
+				}
+
+				if paramFields.Description != "" {
+					sb.WriteString("\n<description>" + strings.TrimSpace(paramFields.Description) + "</description>")
+				}
+
+				if len(paramFields.Enum) > 0 {
+					enumJSON, _ := json.Marshal(paramFields.Enum)
+					sb.WriteString("\n<enum>" + string(enumJSON) + "</enum>")
+				}
+
+				sb.WriteString("\n</parameter>")
+			}
+		}
+
+		if len(fn.Parameters.Required) > 0 {
+			reqJSON, _ := json.Marshal(fn.Parameters.Required)
+			sb.WriteString("\n<required>" + string(reqJSON) + "</required>")
+		}
+
+		sb.WriteString("\n</parameters>")
+		sb.WriteString("\n</function>")
+	}
+
+	sb.WriteString("\n</tools>")
+
+	sb.WriteString("\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+		"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+		"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+		"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+		"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+		"- Required parameters MUST be specified\n" +
+		"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+		"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>")
+
+	return sb.String()
+}
+
+func (r *Nemotron3NanoRenderer) buildContent(message api.Message) string {
+	// The parser always extracts thinking into the Thinking field,
+	// so Content will never have <think> tags embedded
+	if message.Thinking != "" {
+		return "<think>\n" + message.Thinking + "\n</think>\n" + message.Content
+	}
+	return "<think></think>" + message.Content
+}
+
+func (r *Nemotron3NanoRenderer) formatContent(content string, truncate bool, addNewline bool) string {
+	if content == "" {
+		return "<think></think>"
+	}
+
+	if !truncate {
+		if addNewline {
+			return strings.TrimSpace(content) + "\n"
+		}
+		return strings.TrimSpace(content)
+	}
+
+	// Truncate thinking - keep only content after </think>
+	c := content
+	if strings.Contains(c, "</think>") {
+		parts := strings.Split(c, "</think>")
+		c = parts[len(parts)-1]
+	} else if strings.Contains(c, "<think>") {
+		parts := strings.Split(c, "<think>")
+		c = parts[0]
+	}
+	c = "<think></think>" + strings.TrimSpace(c)
+
+	if addNewline && len(c) > len("<think></think>") {
+		return c + "\n"
+	}
+	if c == "<think></think>" {
+		return c
+	}
+	return strings.TrimSpace(c)
+}
+
+func (r *Nemotron3NanoRenderer) writeToolCalls(sb *strings.Builder, toolCalls []api.ToolCall) {
+	for _, tc := range toolCalls {
+		sb.WriteString("<tool_call>\n<function=" + tc.Function.Name + ">\n")
+		for name, value := range tc.Function.Arguments {
+			sb.WriteString("<parameter=" + name + ">\n" + r.formatArgValue(value) + "\n</parameter>\n")
+		}
+		sb.WriteString("</function>\n</tool_call>\n")
+	}
+}
+
+func (r *Nemotron3NanoRenderer) formatArgValue(value any) string {
+	switch v := value.(type) {
+	case map[string]any, []any:
+		jsonBytes, _ := json.Marshal(v)
+		return string(jsonBytes)
+	default:
+		return fmt.Sprintf("%v", v)
+	}
+}

model/renderers/renderer.go

@@ -76,6 +76,12 @@ func rendererForName(name string) Renderer {
 		// Used for Olmo-3-32B-Think
 		renderer := &Olmo3ThinkRenderer{Variant: Olmo3Think32B}
 		return renderer
+	case "nemotron-3-nano":
+		renderer := &Nemotron3NanoRenderer{IsThinking: false}
+		return renderer
+	case "nemotron-3-nano-thinking":
+		renderer := &Nemotron3NanoRenderer{IsThinking: true}
+		return renderer
 	default:
 		return nil
 	}