LFM2.5-24B-A2B
24B MoE with 2.45B active, 131K context, always reasons before answering. Built for fast on-device agents; vendor claims it beats Qwen3.6-35B-A3B on LiveClaw.
Recommended quants
- Unsloth UD-Q4_K_XL (14.9 GB)
unsloth/LFM2.5-24B-A2B-GGUF - UD-Q3_K_XL (10.9 GB) / UD-Q2_K_XL (8.0 GB)
unsloth/LFM2.5-24B-A2B-GGUFFits 12 GB / 8-10 GB cards. - Q8_0 (25.5 GB)
unsloth/LFM2.5-24B-A2B-GGUF - Official GGUF / MLX-4bit
LiquidAI/LFM2.5-24B-A2B-GGUF, LiquidAI/LFM2.5-24B-A2B-MLX-4bit
Chat template Works
Always thinks; no non-thinking mode documented.
Fix: Expect reasoning tokens on every reply; budget output length accordingly.
Sampling settings
| Official | temp 0.1, top_k 50 |
|---|
Runtimes
| vLLM | Unverified | |
|---|---|---|
| llama.cpp | Works | Official guide (docs.liquid.ai). |
| Ollama | Unverified | Library page lfm2.5 exists; 24B tag not checked. |
| LM Studio | Works | Official guide. |
| MLX | Works | Official MLX-4bit repo. |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)