LFM2.5-2.6B: Liquid's On-Device Agentic Open Weights
Liquid AI LFM2.5-2.6B (~Aug 4, 2026): ~2.6B open-weight on-device agentic model, ~128K context, LFM custom license, self-reported IF/tool scores. Weights on Hugging Face.
At a Glance
- Org: Liquid AI
- HF:
LiquidAI/LFM2.5-2.6B,LiquidAI/LFM2.5-2.6B-Base(+ GGUF / ONNX / MLX) - Params: ~2.69B (30 layers; short-conv + GQA hybrid)
- Context: 128K-class (131,072 tokens on card)
- License: LFM custom (confirm LFM Open License terms before commercial ship)
- Runtimes: llama.cpp, MLX, vLLM, SGLang, ONNX, Transformers
- Vendor CPU decode: ~220 tok/s M5 Max; ~113 tok/s Ryzen AI Max+ 395; ~30 tok/s phone class; under ~2.5 GB
Selected vendor scores (self-reported)
| Benchmark | LFM2.5-2.6B | Note |
|---|---|---|
| AIME25 | 51.87 | Vendor table |
| LiveCodeBench v6 | 59.41 | Vendor table |
| IFBench | 59.17 | Vendor table |
| BFCLv4 | 56.88 | Vendor table |
| ToolSandbox | 77.83 | Vendor table |
| Claw-Eval EN avg | 62.85 | Vendor table |
| PinchBench | 68.22 | Vendor table |
| BrowseComp+ (OpenClaw) | 26.89 | Vendor table |
| AA-Omniscience-Public | -29.50 | Bare bench name; vendor-cited |
Compared in-blog to Gemma 4 E2B/E4B and Qwen3.5 4B/9B under Liquid's harness notes. Not LLM Stats verified.
What's New
- Agentic post-train pipeline: SFT → teacher specialists → MOPD → agentic RL inside real harnesses (Hermes, OpenClaw, etc.).
- Explicit on-device pitch: free marginal tokens, privacy, parallel local agents.
- Day-one edge packaging (GGUF / MLX / ONNX) alongside GPU servers.
When to Use It
Good fit: High-volume local agents, tool/IF heavy edge apps, privacy-sensitive loops where 2.6B is enough.
Not automatic: Liquid itself flags agentic coding and knowledge-heavy tasks as weaker fits; larger models still win coding in their table.
Caveats
- License is not Apache/MIT; read LFM terms (commercial / redistribution).
- Always-on reasoning style in the chat template (thinks before answering).
- All headline numbers above are vendor-reported under stated decoding settings.
Sources
Questions
Frequently Asked Questions
- Liquid AI's ~2.6B open-weight agentic model (~Aug 4, 2026) for on-device / edge agents: plan, call tools, and run multi-step tasks without a cloud API. Hugging Face:
LiquidAI/LFM2.5-2.6BandLiquidAI/LFM2.5-2.6B-Base(+ GGUF / ONNX / MLX). - 128K-class context (131,072 tokens on the Hugging Face card).
- Liquid's LFM license (lfm1.0 / custom; not Apache/MIT). Confirm LFM Open License terms before commercial ship or redistribution.
No. The selected scores (AIME25, LiveCodeBench v6, IFBench, BFCLv4, ToolSandbox, and others including AA-Omniscience-Public) are from Liquid's vendor table / self-reported under stated decoding settings. Not LLM Stats verified.
- Good fit: high-volume local agents, tool/IF-heavy edge apps, privacy-sensitive loops where 2.6B is enough. Not automatic: Liquid flags agentic coding and knowledge-heavy tasks as weaker fits; larger models still win coding in their table.
Vendor CPU decode figures: ~220 tok/s on M5 Max; ~113 tok/s on Ryzen AI Max+ 395; ~30 tok/s phone class; under ~2.5 GB.
Continue Reading