feat: add LFM 3 (1.7B, VL 2B, Audio 2B)

apps/nlp/app/llm/index.tsx

@@ -25,6 +25,7 @@ import { Button } from '../../components/Button';
 import { getImage, skImageToBuffer } from '../../utils';
 import {
   LLM_MODELS,
+  parseLfm3ToolCalls,
   parseMiniCpmToolCalls,
   type LLMModelConfig,
   type LLMBackend,
@@ -49,6 +50,13 @@ function formatStats(stats: llm.LLMGenerationStats): string {
   );
 }
 
+// With thinking off, LFM 3 can still write a short draft closed by </think>
+// before its answer; show only the answer.
+function stripDraft(text: string): string {
+  const end = text.lastIndexOf('</think>');
+  return end === -1 ? text : text.slice(end + '</think>'.length).trimStart();
+}
+
 function LLMContent() {
   const insets = useSafeAreaInsets();
   const [selectedBackend, setSelectedBackend] = useState<LLMBackend>('xnnpack');
@@ -103,10 +111,15 @@ function LLMContent() {
   const thinkingRef = useRef(enableThinking);
   const supportsThinkingSwitch = activeModel.supportsThinkingSwitch !== false;
   const toolOpts = useMemo(() => {
-    if (activeModel.toolOpts?.parseToolCalls !== parseMiniCpmToolCalls) return activeModel.toolOpts;
+    // Parsers whose output layout depends on the thinking switch.
+    const opts = activeModel.toolOpts;
+    const thinkingAware = [parseMiniCpmToolCalls, parseLfm3ToolCalls].find(
+      (p) => p === opts?.parseToolCalls
+    );
+    if (!opts || !thinkingAware) return opts;
     return {
-      ...activeModel.toolOpts,
-      parseToolCalls: (text: string) => parseMiniCpmToolCalls(text, thinkingRef.current),
+      ...opts,
+      parseToolCalls: (text: string) => thinkingAware(text, thinkingRef.current),
     };
   }, [activeModel]);
 
@@ -303,6 +316,7 @@ function LLMContent() {
         payload = textMessage;
       }
 
+      const thinkingOff = supportsThinkingSwitch && !enableThinking;
       const result = await sendMessage(
         payload,
         (token) => {
@@ -319,7 +333,8 @@ function LLMContent() {
         )
         .map((m) => ({
           role: m.role,
-          content: typeof m.content === 'string' ? m.content : '',
+          content:
+            typeof m.content !== 'string' ? '' : thinkingOff ? stripDraft(m.content) : m.content,
           stats: m.role === 'assistant' ? result.stats[assistantIdx++] : undefined,
           toolCalls: m.role === 'assistant' ? m.toolCalls : undefined,
         }));
@@ -533,7 +548,11 @@ function LLMContent() {
             {streamingResponse !== null && (
               <View style={styles.turn}>
                 <View style={[styles.bubble, styles.assistantBubble]}>
-                  <Text style={styles.assistantText}>{streamingResponse || '…'}</Text>
+                  <Text style={styles.assistantText}>
+                    {(supportsThinkingSwitch && !enableThinking
+                      ? stripDraft(streamingResponse)
+                      : streamingResponse) || '…'}
+                  </Text>
                 </View>
               </View>
             )}

apps/nlp/constants/llm.ts

@@ -284,6 +284,48 @@ export function parseLfmToolCalls(text: string): llm.ToolParserResult | undefine
   return { toolCalls, textContent: textContent || undefined };
 }
 
+export function parseLfm3ToolCalls(
+  text: string,
+  enableThinking = true
+): llm.ToolParserResult | undefined {
+  // With thinking on, the generation prompt opens <think>: only parse the
+  // final answer. With it off, the prompt already closes it.
+  const thinkingEnd = text.lastIndexOf('</think>');
+  if (thinkingEnd === -1 && (enableThinking || text.includes('<think>'))) return undefined;
+  const answer = thinkingEnd === -1 ? text : text.slice(thinkingEnd + '</think>'.length);
+  // <tool_call>\n<function=NAME>\n<parameter=KEY>\nVALUE\n</parameter>\n</function>\n</tool_call>
+  const callRegex = /<tool_call>\s*<function=([^>\n]+)>([\s\S]*?)<\/function>\s*<\/tool_call>/g;
+  const toolCalls: llm.ToolCall[] = [];
+
+  for (const match of answer.matchAll(callRegex)) {
+    const name = match[1]!.trim();
+    const tool = TOOLS.find((candidate) => candidate.function.name === name);
+    if (!tool) return undefined;
+    const params = tool.function.parameters;
+    const paramRegex = /<parameter=([^>\n]+)>\n?([\s\S]*?)\n?<\/parameter>/g;
+    const args: Record<string, unknown> = {};
+    try {
+      for (const param of match[2]!.matchAll(paramRegex)) {
+        const key = param[1]!.trim();
+        const schema = params?.properties?.[key] as { type?: string } | undefined;
+        if (!schema || Object.prototype.hasOwnProperty.call(args, key)) return undefined;
+        // The template renders strings verbatim and everything else as JSON.
+        args[key] = schema.type === 'string' ? param[2]! : JSON.parse(param[2]!);
+      }
+    } catch {
+      return undefined;
+    }
+    toolCalls.push({ type: 'function', function: { name, arguments: args } });
+  }
+
+  if (toolCalls.length === 0) return undefined;
+  const textContent = answer
+    .replace(callRegex, '')
+    .replace(/<\|im_end\|>/g, '')
+    .trim();
+  return { toolCalls, textContent: textContent || undefined };
+}
+
 function decodeMiniCpmArgument(value: string): string {
   const trimmed = value.trim();
   if (trimmed.startsWith('<![CDATA[') && trimmed.endsWith(']]>')) {
@@ -609,11 +651,19 @@ export const BIELIK_V3_GENERATION_CONFIG: llm.LLMGenerationConfig = {
 
 export const DEFAULT_SYSTEM_PROMPT = 'You are a helpful assistant. Your name is HAL-9000';
 
+// LFM 3 Audio model card prompt for transcription.
+export const LFM3_AUDIO_SYSTEM_PROMPT = 'Perform ASR in English.';
+
 export const LFM_TOOL_OPTS: LLMToolOpts = {
   tools: TOOLS,
   parseToolCalls: parseLfmToolCalls,
 };
 
+export const LFM3_TOOL_OPTS: LLMToolOpts = {
+  tools: TOOLS,
+  parseToolCalls: parseLfm3ToolCalls,
+};
+
 export const MINICPM5_TOOL_OPTS: LLMToolOpts = {
   tools: TOOLS,
   parseToolCalls: parseMiniCpmToolCalls,
@@ -639,6 +689,18 @@ export const LLAMA_TOOL_OPTS: LLMToolOpts = {
   parseToolCalls: parseLlamaToolCalls,
 };
 
+/**
+ * LFM3 sampling defaults from the upstream generation_config.json.
+ * @see https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/generation_config.json
+ */
+export const LFM3_GENERATION_CONFIG: llm.LLMGenerationConfig = {
+  temperature: 0.1,
+  topP: 1,
+  topK: 50,
+  repetitionPenalty: 1.1,
+  maxNewTokens: 512,
+};
+
 // --- Model Config Type ---
 
 export type LLMBackend = 'xnnpack' | 'mlx' | 'vulkan';
@@ -680,6 +742,72 @@ export const LLM_MODELS: LLMModelConfig[] = [
     iosOnly: true,
   },
 
+  // LFM 3 family
+  {
+    id: 'lfm-3-1.7b-xnnpack',
+    name: 'LFM 3 1.7B',
+    backend: 'xnnpack',
+    model: models.llm.LFM3_1_7B.XNNPACK_8DA4W,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: DEFAULT_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: LFM3_TOOL_OPTS,
+  },
+  {
+    id: 'lfm-3-1.7b-mlx',
+    name: 'LFM 3 1.7B',
+    backend: 'mlx',
+    model: models.llm.LFM3_1_7B.MLX_INT4,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: DEFAULT_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: LFM3_TOOL_OPTS,
+    iosOnly: true,
+  },
+  {
+    id: 'lfm-3-vl-2b-xnnpack',
+    name: 'LFM 3 VL 2B',
+    backend: 'xnnpack',
+    model: models.llm.LFM3_VL_2B.XNNPACK_8DA4W,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: DEFAULT_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: undefined,
+  },
+  {
+    id: 'lfm-3-vl-2b-mlx',
+    name: 'LFM 3 VL 2B',
+    backend: 'mlx',
+    model: models.llm.LFM3_VL_2B.MLX_INT4,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: DEFAULT_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: undefined,
+    iosOnly: true,
+  },
+
+  {
+    id: 'lfm-3-audio-2b-xnnpack',
+    name: 'LFM 3 Audio 2B',
+    backend: 'xnnpack',
+    model: models.llm.LFM3_AUDIO_2B.XNNPACK_8DA4W,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: LFM3_AUDIO_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: undefined,
+  },
+  {
+    id: 'lfm-3-audio-2b-mlx',
+    name: 'LFM 3 Audio 2B',
+    backend: 'mlx',
+    model: models.llm.LFM3_AUDIO_2B.MLX_INT4,
+    generationConfig: LFM3_GENERATION_CONFIG,
+    systemPrompt: LFM3_AUDIO_SYSTEM_PROMPT,
+    stopRegex: undefined,
+    toolOpts: undefined,
+    iosOnly: true,
+  },
+
   // LFM 2.5 family
   {
     id: 'lfm-2.5-2.6b-xnnpack',