LFM3 on the Mac
Verdict
LFM3 can call tools. It is weak at long agentic coding, and so is every other 2B model. Use it as a fast local worker inside Allternit, not as the coding brain.
Tool calling vs agentic coding
| Task | Steps | LFM3 at 2B |
|---|---|---|
| Call one tool with correct arguments | 1 to 3 | Yes. Trained on tool calls and agentic RL |
| Browse or operate an app | 5 to 20 | Yes, with limits. WebArena-Verified 38.4 (vendor score) |
| Agentic coding in a repo | 50 or more, large context | No. The model card says "not recommended" |
The models you named are larger. Gemma 4 12B and Bonsai 9B or 27B are 4 to 15 times the size. A 2B Qwen or Gemma has the same coding limit as LFM3.
What we can do with them
| Use | Model | Why it fits |
|---|---|---|
| Gizzi sidecar: titles, compaction, short tool calls | LFM3-1.7B | 1 GB, 85 tok/s, replaces the 4B sidecar |
| Read the screen for computer use | LFM3-VL-2B | Local, no cloud call for each screenshot |
| Offline voice with tool calls | LFM3-Audio-2B | Speech in and out, 18 languages |
| Route requests and gate actions | d1-2B | Typed answer plus a probability in one pass |
| Phone brain in the PWA | LFM3-1.7B ONNX | Bonsai had no ONNX build. LFM3 has one. Browser use not tested |
Measured speed (M1 Pro, 32 GB, Q4_K_M, Metal)
| Model | Read prompt (512) | Read prompt (8192) | Write |
|---|---|---|---|
| LFM3-1.7B | 1169 tok/s | 1097 tok/s | 85 tok/s |
| LFM3-VL-2B (text part) | 1169 tok/s | 1097 tok/s | 88 tok/s |
Caveats
- Scores come from Liquid AI. No independent tests yet.
- Homebrew llama.cpp, mlx-lm and Ollama do not run LFM3 yet.
- License: free for commercial use only while annual revenue is under $10M.