LFM3 launched on 27 October 2026 with a 256K-token text model small enough to run on a laptop. It suits teams running local agents and fine-tuned on-device models, provided their annual revenue stays under the 10 million dollar license threshold.
Can you use it?
- Liquid AI API: yes (Hosted API; LFM3 per-token rates not listed on the pricing page).
- Liquid Playground: yes. Plans and limits not published on the pages read.
- Free plan: yes.
- Open weights: released (LFM Open License v1.0).
Checked 2 Nov against Liquid AI's own pages. Compare access across frontier models
What changed
- · Release LFM3 family released: six open-weight models near 2B parameters source
LFM3 is Liquid AI's October 2026 family of six open-weight models of roughly 2B parameters each, covering text, vision, audio, omni and decision tasks from one backbone. The text model, LFM3-1.7B, reaches 43.40 on SWE-bench Verified in Liquid's own OpenCode runs.
Where it sits
- 43.4%% solvedSWE-bench VerifiedHigher is better#35 / 37peer median 76.5%per source, see benchmark scores
- 63.4%% correctGPQA DiamondHigher is better#53 / 60peer median 87.6%per source, see benchmark scores
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Provider: Liquid AI · Family: LFM3
Context window: 256,000 tokens
Input modalities: text, image, video, audio · Output: text, audio
About LFM3
LFM3 is the third generation of Liquid Foundation Models from Liquid AI, announced on 27 October 2026 as six open-weight models around 2B parameters that all branch from one LFM3-1.7B-Base checkpoint. The base was pre-trained on 22T tokens, and its last 3.5T tokens are curated mid-training data with agentic traces and tool calls that also stretch the context window to 256K tokens. The architecture mixes 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers across 24 layers. Liquid says the 1.7B text model beats its earlier LFM2.5-2.6B while using 35% fewer parameters.
The six models, read from their Hugging Face cards on 2 November 2026:
LFM3-1.7B: text in, text out; 1.73B parameters; 256K-token context. LFM3-VL-2B: text, images and video in, text out; 1.92B parameters; 128K-token context. LFM3-Audio-2B: text or audio in, text or speech out; 2B parameters; 32K-token context. LFM3-Omni-2B: text, images, video and audio in, text or speech out; 2.17B parameters; 32K-token context. d1-2B: text, images and video in, calibrated answers out; 1.92B parameters; 128K-token context. d1-Omni-2B: text, images and audio in, calibrated answers out; 2.17B parameters; 128K-token context.
Every card lists the same license, the LFM Open License v1.0, and none of the six cards says Apache 2.0, although some secondary coverage claims it does. The license text grants a no-charge copyright and patent license but conditions commercial use on staying under a threshold of 10 million US dollars in annual revenue; a legal entity above that threshold is not licensed for commercial use under the agreement and needs a separate deal with Liquid. Qualified non-profits using it for non-commercial or research purposes are exempt.
Liquid reports results for the 1.7B text model against LFM2.5-2.6B, Gemma 4 E4B and two Qwen3.5 sizes: 43.40 on SWE-bench Verified with the OpenCode setup, 63.37 on GPQA Diamond and 73.46 on LiveCodeBench v6. Larger models still win on agentic coding and knowledge, and Liquid says as much. The vision model is compared with Gemini 3.5 Flash Lite, GPT-5.6 Luna and Gemma 4 12B across 62 vision benchmarks, and Liquid reports an average of 71.4 for LFM3-VL-2B against 73.4 for GPT-5.6 Luna. All of these are vendor-reported, and neither the announcement nor the cards include a safety evaluation. It succeeds the vision line of LFM2.5-VL-3B, though Liquid does not call LFM3 a formal replacement.
The models run through llama.cpp, MLX, vLLM, SGLang, Ollama, Transformers and ONNX, with GGUF, FP8, NVFP4 and MLX builds on Hugging Face; the text model fits in about 1 GB at 4-bit precision. Local tools such as Ollama, LM Studio and llama.cpp are the usual route, and hosted access comes from Liquid's own Playground and API, where Liquid's pricing page lists no LFM3 rates and sends buyers to sales. OpenRouter is named as a day-0 launch partner.
Pricing
Weights are free to download under the LFM Open License v1.0, so cost is your own hardware. No LFM3 API rate is published, so Hokai leaves the API prices empty. See [Liquid AI](/hub/companies/liquid-ai) for the company and [LFM2.5-VL-3B](/hub/models/lfm2.5-vl-3b) for the previous vision model.
Key Features
- One backbone, six models: Text, vision, audio, omni and two decision models branch from LFM3-1.7B-Base, so you pick the one that fits the job.
- 256K-token text model: LFM3-1.7B is post-trained for agent frameworks such as Hermes Agent, Pi and OpenClaw.
- Video and speech: LFM3-VL-2B accepts video, and LFM3-Audio-2B handles speech in 18 languages.
- Decision models: d1-2B and d1-Omni-2B answer many typed questions at once with calibrated probabilities.
Pros
- Runs on phones and laptops, with the text model near 1 GB at 4-bit precision.
- Strong tool-use scores for its size in Liquid's own comparison tables.
- Wide runtime support, from llama.cpp and MLX to vLLM and SGLang.
Cons
- Revenue-gated commercial license, not a standard open-source license.
- No safety evaluation published with the release.
- Weaker than larger models on coding and knowledge benchmarks.
Benchmarks
- IFBench: 73.0% vendor-reported · 02 Nov 2026 — How precisely the model follows detailed instructions, % passing.
- MCP Atlas: 32.6% vendor-reported · 02 Nov 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
- GPQA Diamond: 63.4% vendor-reported · 02 Nov 2026 — PhD-level science questions that are hard to search for, % correct.
- LiveCodeBench v6: 73.5% vendor-reported · 02 Nov 2026 — Recently published coding problems, % solved.
- SWE-bench Verified: 43.4% vendor-reported · 02 Nov 2026 — Real GitHub issues fixed end to end, % solved.
- Terminal-Bench 2.1: 12.9% vendor-reported · 02 Nov 2026 — Multi-step tasks completed in a real command line, % solved.
- Humanity's Last Exam: 7.2% vendor-reported · 02 Nov 2026 — Expert-written questions across many fields, % correct.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does it cost to run LFM3?
The weights are free to download, so your cost is hardware. Liquid's pricing page lists no per-token rate for LFM3 and points buyers to a sales contact, so Hokai leaves the API price empty. Companies above the license revenue threshold need a paid agreement for commercial use.
Which LFM3 license applies to companies?
All six models use the LFM Open License v1.0, which Hokai read on the LFM3-1.7B repository. Commercial use is licensed only while annual revenue stays under 10 million US dollars, and larger companies must contact Liquid. Treat claims of Apache 2.0 on the audio or vision models as wrong until a card says so.
How does LFM3-1.7B compare with Qwen3.5-9B on agent tasks?
In Liquid's table the 1.7B model leads on tool-use scores such as BFCL V4 (71.39 against 62.83) and beats Qwen3.5-9B on web and business-app tasks. Qwen3.5-9B stays ahead on SWE-bench Verified (51.20) and GPQA Diamond (80.40). Choose LFM3 for small, fast local agents and the larger model for knowledge-heavy work.
Can LFM3 handle audio and video?
Yes, through different models. LFM3-VL-2B takes video, LFM3-Audio-2B handles speech in 18 languages, and LFM3-Omni-2B combines text, images, video and audio. The plain 1.7B text model takes text only.
Who should avoid LFM3?
Skip it if your company exceeds the revenue threshold and has no Liquid agreement, or if you need a documented safety evaluation, since neither the announcement nor the cards include one. It is also the wrong pick for frontier-level knowledge or coding tasks, where Liquid's own tables show larger models ahead.