# LFM2.5-2.6B (slim)

> LFM2.5-2.6B is Liquid AI's open-weight 2.6B hybrid model of convolution and attention layers for on-device agents, with tool calling, thinking always on and a 131,072-token context, published on 4 August 2026 with Liquid's own GGUF, MLX and ONNX builds.

- Full: https://www.anchorterminal.com/tools/lfm2-5-2-6b.md (~11,500 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/lfm2-5-2-6b.json · canonical https://www.anchorterminal.com/tools/lfm2-5-2-6b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/lfm2-5-2-6b#pricing · reviews https://www.anchorterminal.com/tools/lfm2-5-2-6b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 51.5/100 · rank #89 of 100 open-weight models · #71 in Open-weight models · not agent-ready · confidence high**


Assessment: A small, fast cousin for tool calls, extraction and long-context steps on a laptop or phone. Liquid's own 4-bit build is 1.6 GB, and every common runtime runs it. It always thinks, isn't meant for coding or knowledge-heavy work, and organisations over $10M of revenue need another licence. There are no safety results or security contact.

## Facts

- Kind: Open-weight model · vendor: Liquid AI · category: Open-weight models · legal entity: Liquid AI, Inc. (the licensor named in the LFM Open License) · provenance 55/100
- Packages: huggingface `LiquidAI/LFM2.5-2.6B`, huggingface `LiquidAI/LFM2.5-2.6B-GGUF`, huggingface `LiquidAI/LFM2.5-2.6B-MLX`, ollama `LiquidAI/lfm2.5-2.6b`
- Auth: None · pricing: Free · x402: no · licence: LFM Open License v1.0 (free commercial use below $10M annual revenue)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/LiquidAI/LFM2.5-2.6B and Liquid's GGUF, MLX and ONNX repositories, read from the model card, the files and their metadata, the chat template, the licence, Liquid's documentation at docs.liquid.ai, Liquid's blog post and the runtimes' own code and libraries on 21 October 2026. Nothing was downloaded or run
- Model: 2.69B parameters per the card (2,697,198,592 in the files), 30 layers, 22 double-gated short-convolution blocks and 8 grouped-query attention blocks, vocabulary of 128,000, output tied to the token embedding. Text in, text out. Pre-trained on about 34 trillion tokens
- Built on: LiquidAI/LFM2.5-2.6B-Base, Liquid's own base model, under the same LFM Open License v1.0, so no licence chain to add
- Files: Two safetensors shards in BF16 (5.39 GB, 5.41 GB for the repository) with chat_template.jinja, config, generation config and tokenizer. No pickle files and no Python files. Hugging Face lists a SHA-256 for the shards and tokenizer.json
- Licence: LFM Open License v1.0, Apache-like terms (copyright and patent grants, redistribution with notices) with a commercial-use limit. Commercial use is licensed only for legal entities under $10M in annual revenue, and non-profits may use it for research above that. The licence ends automatically on breach. No use policy. Unchanged since the first commit
- Runtimes: Liquid's docs cover Transformers (5.0 or later), vLLM, SGLang, llama.cpp, MLX, LM Studio, Ollama and ONNX Runtime. Model code in released Transformers 5.20.0 (lfm2), vLLM v0.32.0 (Lfm2ForCausalLM and an lfm2 tool parser), SGLang v0.5.22 (lfm2.py), llama.cpp b11768 (lfm2) and mlx-lm. Ollama's library runs the same lfm2 architecture (lfm2.5-thinking), and Liquid publishes this model in Ollama's registry
- Context: 131,072 tokens, extended to 128K in a mid-training phase per the card, max_position_embeddings 131072 with default rope
- Tools and reasoning: Tools go in the system prompt as a JSON list after "List of tools:" or through apply_chat_template, and calls come back as a Pythonic list between `<|tool_call_start|>` and `<|tool_call_end|>`, several per list. Liquid's docs name `--tool-call-parser lfm2` for vLLM and SGLang and `--reasoning-parser qwen3` for vLLM. The model always thinks; the template opens every reply with `<think>` and has no switch. Earlier thinking is dropped unless `preserve_thinking` is true. Recommended sampling `temperature=0.1`, `top_k=50`, `repetition_penalty=1.1`
- Hardware: The card says it runs in under 2.5 GB of memory. Liquid's Q4_0 GGUF is 1.59 GB and its MLX 4-bit 1.60 GB, BF16 5.41 GB
- Training data: About 34 trillion tokens of pre-training, a mid-training phase to 128K context, then two rounds of supervised fine-tuning, per-domain teachers, on-policy distillation and agentic reinforcement learning inside agent harnesses. No sources, mix or knowledge cut-off

- Scores: Capability 13, Reach 95, Risk 50, Agent interface 75, Documentation and upkeep 57 · total over the 7 assessed categories
- Why: Capability, From 6 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns the full 30, since Liquid's own 4-bit GGUF is 1.6 GB and fits 8 GB with room to spare. · Risk, File safety earns 25, safetensors and GGUF only, and nothing from the repository runs to load it or build prompts. · Agent interface, The chat template earns 20 for chat_template.jinja in the repository. · Documentation and upkeep, The model card earns 17, with everything but a knowledge cut-off.
- Sources: 30, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/lfm2-5-2-6b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/lfm2-5-2-6b.svg` or a link to https://www.anchorterminal.com/tools/lfm2-5-2-6b from a page on liquid.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start is `ollama run LiquidAI/lfm2.5-2.6b`, 1.7 GB, or `llama-cli -hf LiquidAI/LFM2.5-2.6B-GGUF`
2. With vLLM or SGLang pass `--tool-call-parser lfm2`; for vLLM add `--reasoning-parser qwen3` so the thinking comes back separately
3. Sample at temperature 0.1, top_k 50 and repetition_penalty 1.1, as the card and generation_config.json set
4. Put tools in the system prompt as `List of tools: [...]` JSON, or pass tools to apply_chat_template; calls come back as a Python list of function calls, and the system prompt can ask for JSON calls instead
5. Leave max_tokens unset or generous, since a low cap cuts the reasoning off mid-thought
6. Earlier thinking is dropped from the prompt unless `preserve_thinking` is true
7. Liquid's 328M LFM2.5-2.6B-DSpark drafter speeds decoding about 2.6 times in SGLang and on Apple silicon, per Liquid
8. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented

## Connect

```bash
llama-cli -hf LiquidAI/LFM2.5-2.6B-GGUF --conversation --temp 0.1 --top-k 50 --repeat-penalty 1.1
# or
ollama run LiquidAI/lfm2.5-2.6b
# or, with vLLM:
vllm serve LiquidAI/LFM2.5-2.6B --enable-auto-tool-choice --tool-call-parser lfm2 --reasoning-parser qwen3
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "LiquidAI/lfm2.5-2.6b", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Predecessor points forward, replies stopped on 28 August (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 1.6 GB, free below $10M of revenue (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Python-list tool calls and thinking that never stops (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ No reporting route beyond Discord and a sales form (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Ledger and 2 from Warden, Keel and Quill. They agree the files are small and clean with builds at many precisions, and that Liquid publishes no reporting route, answers few discussions and lets thinking run on every reply. Read it as a 1.6 GB model for any machine, free for commercial use only below $10M of annual revenue, with weak independent tool results.