Liquid AI's LFM2.5-24B-A2B runs local agents on 2.45B active parameters — and beats Qwen3.6-35B-A3B on tool calling but not on SWE-Bench
Liquid AI released LFM2.5-24B-A2B on October 13, 2026: an open-weight 24B MoE reasoning model with ~2B active parameters, 131K context and day-one GGUF, MLX, vLLM and SGLang support. Which GGUF to download, how to point Pi, Hermes or OpenClaw at it, and where it loses to Qwen3.6-35B-A3B.
Liquid AI released LFM2.5-24B-A2B on October 13, 2026 — a 24B-parameter mixture-of-experts reasoning model that activates only about 2B parameters per token, aimed squarely at agents that run on your own machine. Liquid calls it the largest and final model in the LFM2.5 family. (Source: Liquid AI, 2026-10-13)
Key facts:
- 24.0B total, 2.45B active parameters. 40 layers (30 double-gated short-convolution blocks + 10 GQA), 64 experts per MoE block with top-4 routing. (Source: model card)
- Context grew from 32,768 to 131,072 tokens, and the vocabulary doubled from 65K to 128K for better non-Latin tokenization, versus LFM2-24B-A2B.
- Reasoning-only. Unlike its predecessor, it always emits a chain of thought before answering.
- Open weights under Liquid’s own
lfm1.0license — not Apache 2.0, so read it before shipping commercially. - Day-one formats: BF16, FP8, GGUF, ONNX and MLX 4-bit. A DSpark speculative-decoding drafter for SGLang is listed as “coming soon”.
- Hosted API: $0.10 per 1M input tokens ($0.05 cached) and $1.00 per 1M output tokens on Liquid Console.
Where it wins and where it loses
Liquid publishes the full 16-benchmark table, losses included. Against Qwen3.6-35B-A3B — a model with roughly 50% more total and active parameters:
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B |
|---|---|---|
| AutomationBench | 27.69 | 8.72 |
| LiveClaw | 48.09 | 38.19 |
| ToolSandbox | 80.71 | 75.02 |
| IFBench | 74.69 | 64.21 |
| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 |
| WebArena Verified | 26.75 | 37.62 |
| GPQA Diamond | 74.42 | 83.69 |
(Source: LFM2.5-24B-A2B model card, all models evaluated by Liquid with vLLM.)
The pattern matches what we saw with Meta’s Muse Glimmer 30B: strong at choosing and formatting tool calls, weaker at actually writing the patch. Liquid says as much — it recommends pairing the model with retrieval or web search for knowledge-heavy work. If your local agent is mainly a coding agent, the 15-point SWE-Bench gap is the number to weigh; our Qwen 3.6 local coding setup remains the stronger default there.
Which GGUF, and how fast
The official GGUF repository lists these files (sizes from the Hugging Face file tree, retrieved 2026-10-15):
| File | Size |
|---|---|
LFM2.5-24B-A2B-QAD-Q4_0.gguf | 13.6 GB |
LFM2.5-24B-A2B-Q4_K_M.gguf | 14.5 GB |
LFM2.5-24B-A2B-Q5_K_M.gguf | 17.0 GB |
LFM2.5-24B-A2B-Q6_K.gguf | 19.7 GB |
LFM2.5-24B-A2B-Q8_0.gguf | 25.5 GB |
| BF16 / F16 | 48.0 GB |
The Q4_0 file is the interesting one: it is a Quantization-Aware Distillation checkpoint, trained to recover accuracy lost to Q4_0, and it is the smallest download. Liquid’s model-card command:
llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF \
--hf-file LFM2.5-24B-A2B-QAD-Q4_0.gguf \
--temp 0.1 --top-k 50 --repeat-penalty 1.05Those sampling values are also stored in each GGUF’s general.sampling.* metadata, so llama.cpp applies them by default. Liquid’s headline decode numbers — 184 tok/s on an Apple M5 Max and 114 tok/s on an AMD Ryzen AI Max+ 395 — were measured with llama.cpp at Q4_K_M with a 4K-token input context, not at full 131K context. Expect lower figures on long agent transcripts. (Source: GGUF repository; model card)
Pointing an agent harness at it
The model was RL-trained inside OpenClaw, Hermes Agent, Pi and OpenCode, and the model card gives configs for the first three against an OpenAI-compatible endpoint on http://localhost:8080/v1 — llama-server’s default port in Liquid’s llama.cpp docs. For Hermes:
hermes config set model.provider custom
hermes config set model.base_url http://localhost:8080/v1
hermes config set model.default LFM2.5-24B-A2B
hermes config set model.context_length 131072
hermes config set model.api_mode chat_completions
hermes config set agent.tool_use_enforcement trueOne gotcha: by default LFM2.5 writes Pythonic tool calls between <|tool_call_start|> and <|tool_call_end|> tokens, not JSON. If a harness fails to parse calls, the model card says you can ask for JSON function calls in the system prompt. If you plan to share one llama-server among several agents, see how to scale llama-server for multiple users before raising the context to 131K.
Sources
Source: Liquid AI