Liquid AI LFM3 Review: Six 2B Open-Weight Models, Free Under $10M

The short version

Liquid AI launched LFM3 on October 27, 2026: six open-weight models around 2 billion parameters, all branched from one shared backbone. There is a text model, a vision model, an audio model, an omni model, and two decision models that answer questions instead of writing text. Every one of them is free to download, fine-tune, and use commercially if your company earns under $10 million a year. The catch is on the API side: Liquid's own pricing page still lists no per-token rate for any LFM3 model, and OpenRouter has no LFM3 listing yet, so the only way to run LFM3 today with a known price is to host it yourself.

That is a strange split for a launch this polished. The benchmark tables Liquid published are genuinely strong for the size class: the 1.7B text model beats a 9B Qwen on agent benchmarks, the vision model matches a 12B Gemma, and the decision models answer in milliseconds with no output tokens at all. But every one of those numbers is company-run, and the places where LFM3 loses — general knowledge, long-context retrieval, audio tool-calling — are exactly the places a buyer would notice first.

Six models, one backbone

The idea behind LFM3 is specialization by branch, not scale. Instead of one big general-purpose model, Liquid trained a single base model, LFM3-1.7B-Base, on 22 trillion tokens, then branched it into six models that share the architecture and most of the weights. You pick the one that fits the job, fine-tune it on your own data, and run it where your application lives — a laptop, a phone, a Raspberry Pi 5, or a car.

The architecture is a tri-hybrid mix: of 24 layers, 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window, and 6 are global attention. Liquid says the design was found by automated hardware-in-the-loop search across 14 device and backend combinations, from a Samsung Galaxy S26 Ultra and a OnePlus 15 to a Raspberry Pi 5, an Apple M4 Max, AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.

The text model: small, but built for agents

LFM3-1.7B is the headline model, and Liquid positions it as an agent model first and a chat model second. It is post-trained to run inside agent harnesses such as Hermes Agent, Pi, and OpenClaw, and the launch post describes it triaging an inbox, managing a calendar, and researching topics on the web. The training recipe is four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning across math, code, SQL, and structured outputs, distillation from six specialist teacher models, and agentic reinforcement learning inside real harnesses.

On Liquid's own benchmark table, the 1.7B model averages 49.37 across the listed evaluations, ahead of Qwen3.5-9B at 47.24, Qwen3.5-4B at 39.27, MiniCPM5-2B at 39.19, and Gemma 4 E4B at 31.74. The wins are concentrated where Liquid aimed: LiveClaw at 45.60 against 35.77 for the 9B Qwen, BFCL V4 tool-calling at 71.39 against 62.83, WebArena-Verified at 38.42 against 28.94, and AutomationBench at 41.29 against 15.10. It also leads the table on instruction following, with IFEval at 96.66, and posts the highest non-hallucination rate on the public AA-Omniscience test at 92.93 percent.

The losses are just as clear, and Liquid prints them in the same table. On GPQA Diamond, a hard science-knowledge test, LFM3-1.7B scores 63.37 against 75.76 for the 4B Qwen and 80.40 for the 9B Qwen. On Humanity's Last Exam without tools it scores 7.17 against 16.41 for the 9B Qwen, and search only lifts it to 13.90. On RULER at 128K tokens it scores 65.56 against 84.63 and 87.11 for the two Qwen models. On SWE-bench Verified it scores 43.40, behind the 9B Qwen at 51.20. The pattern is consistent: this is a strong small model for agentic work and tool use, and a weak one for raw knowledge and long-context retrieval compared with bigger models. All of these figures are company-run, served with vLLM in BF16, and there is no independent head-to-head yet.

Where the size pays off is speed and memory. Liquid reports that on-device, measured with llama.cpp at Q4_0 and a 32K-token context, LFM3-1.7B decodes 1.8 to 2.8 times faster than the next model on all three devices tested, and uses the least memory. A full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. On a single NVIDIA B300 with SGLang in BF16, it produces 3.6 billion tokens per day at 1,024 concurrent requests. As a demonstration, Liquid ran more than 300,000 LFM3-1.7B agents in parallel to investigate the Enron email archive.

Vision: the strongest card in the family

LFM3-VL-2B is the first LFM vision model to take video as input, and it is the model with the most impressive numbers in the launch. Across 62 vision benchmarks it averages 71.4, against 57.6 for the previous LFM2.5-VL-3B, 64.0 for Qwen 3.5 2B, 53.6 for Gemma 4 E2B, 71.7 for Gemma 4 12B, 73.4 for GPT-5.6 Luna, and 70.0 for Gemini 3.5 Flash Lite. In other words, a roughly 2B-parameter model matches a 12B Gemma, beats Google's Flash Lite, and sits within 2 points of OpenAI's GPT-5.6 Luna — all on Liquid's own benchmark suite, which is the caveat that applies to every number in this paragraph.

The category breakdown shows where it wins and where it does not. It leads the table on GUI understanding at 83.1, visual grounding at 77.3, embodied reasoning at 64.4, and document, OCR, and chart work at 83.9. It trails on video at 61.2 against 72.8 for GPT-5.6 Luna, and on agentic vision tasks at 42.5 against 56.1 for Luna. So the honest read is that this is a very capable small vision model for documents, interfaces, and grounding, and a middling one for video and multi-step visual reasoning.

The vision encoder is new in-house work: a 164M-parameter encoder that Liquid says matches Qwen3.5-2B's vision encoder at 55.9 percent average accuracy across 13 vision-language benchmarks while using 2.2 times fewer FLOPs, and uses 3 times less compute than the SigLIP2 encoder in previous LFM vision models. For video, it jointly embeds each pair of sampled frames into one temporal slice, halving the temporal sequence length. On-device, Liquid claims the highest prefill throughput and lowest memory footprint in its size class.

Audio: good voice, weaker tool-calling

LFM3-Audio-2B takes text or audio in and answers with text, speech, or both, across 18 languages. It adds audio understanding, general-purpose audio reasoning, and tool calling over the previous LFM2.5-Audio. The new audio encoder is fully causal, so it can process live microphone input while you speak, with time to the first token as low as 50 milliseconds. LiquidCodec, the new audio tokenizer, turns generated tokens into speech at about 87 times real time on a single CPU thread on an Apple M4 Max.

The numbers are mixed, and two of them cut against Liquid. On English speech recognition it scores a 6.2 percent word error rate, slightly worse than the older LFM2.5-Audio at 6.1 percent, and behind Voxtral Mini 3B at 5.5 and Kimi-Audio-7B at 5.3. On multilingual recognition it scores 6.0 percent, the best in the table. On VoiceBench it scores 73.1, up from 55.2 for LFM2.5-Audio, but behind MiMo-Audio-7B at 77.3. On speech synthesis it leads on both measures, with a UTMOSv2 quality score of 3.8 and a 1.4 percent word error rate. But on tool-calling — the capability that matters most for a voice assistant that actually does things — it scores 58.5 on Liquid's internal evaluation, behind Gemma 4 E4B at 72.4. On audio understanding it scores 50.6, behind MiMo-Audio-7B at 62.3. The audio model is a genuine step forward for speech quality, and not yet a leader for voice agents that need to act. For the closest voice model with a published price, Mistral's Voxtral 3 runs at $0.04 a minute.

Omni: one model that sees, listens, and talks

LFM3-Omni-2B is Liquid's first omni model, combining vision, audio, and language in a single 2.2B-parameter model. It can take a spoken question about an image, watch a video with its audio track and reason about both, and turn spoken instructions plus visual context into structured tool calls. On Daily-Omni, a video-plus-audio understanding test, it scores 70.1 against 61.3 for Qwen2.5-Omni-7B. On WorldSense it scores 47.1, ahead of both Qwen models. On spoken visual question answering it is competitive but not leading: RealWorldQA at 64.7 against 69.9 for Qwen2.5-Omni-7B, MMStar at 51.8 against 57.3. On OmniBench overall it scores 46.3, behind Qwen2.5-Omni-7B at 49.7.

The use cases Liquid names are a tutor that looks at handwritten math and answers a spoken question, a tablet you walk around with asking about what it sees, and in-car assistants that combine camera and voice. Mercedes-Benz is the named partner here: the company says it is exploring on-device AI with Liquid for its MBUX Virtual Assistant.

Decision models: the answer is a probability, not a paragraph

The two decision models are the most unusual part of LFM3, and the easiest to misunderstand. d1-2B, built on LFM3-VL-2B, reads text and pictures. d1-omni-2B, built on LFM3-Omni-2B, also listens. Neither writes an answer. You give them a situation — a support ticket, an invoice, an app screen, a photo, a voice note — and a list of questions, and they answer all of them in a single forward pass, each with a calibrated probability. That means zero output tokens, which is why the latency is measured in milliseconds rather than seconds.

The new part of this release is typed answers. Beyond yes/no, choices, and scores, both models can now point (draw a box around what you describe, or find every instance of something, with a count), read (pull text, numbers, and dates out of a document, with the exact span they came from, and say "not found" when a field is missing), estimate (give a number with its likely range), and pick (select every option that applies, each with its own probability).

d1-2B replaces the previous d1-3B with about a third fewer parameters, and Liquid reports it beats its predecessor on every decision benchmark listed: Decision Bench on real enterprise records at 85.2 against 83.6, the public Decision Index suite at 51.1 against 49.0, ImajevBench on decisions from photos and rules at 75.5 against 64.0, and a 117-benchmark meta index at 63.2 against 58.8. d1-omni-2B keeps nearly all of its accuracy when questions arrive by voice instead of in writing: DecisionBench at 78.2 spoken against 83.4 written, and JevBench actually higher spoken at 89.0 against 87.6. It also recognizes everyday sounds at 84.5 percent on ESC-50 and identifies which of 43 languages is being spoken at 92.5 percent on FLEURS. All of these are company-reported figures with no independent verification yet, and the Decision Index is Liquid's own benchmark, which is worth knowing before quoting the scores.

What it actually costs

This is where the launch gets awkward, because the honest answer is that nobody has published a price for running LFM3 through an API. Liquid's pricing page, read on October 28, 2026, lists exactly three models with per-token rates: d1 at $0.02 per million input tokens and $0.00 for output, d1-ultra at $0.042 per million input tokens and $0.00 for output, and LFM2.5-24B-A2B at $0.05 input, $0.025 cached input, and $0.70 output per million tokens. There is no LFM3 row on that page. OpenRouter's public model list, checked the same day, carries Liquid's LFM2.5 models but no LFM3 model at all, even though OpenRouter is named as a day-0 launch partner.

What is published, and it matters more than the missing API rate, is the license. The LFM Open License v1.0 makes the open-weight models free to use, including commercially, for companies with under $10 million in annual revenue. Above that line you need a paid enterprise license, priced by deployment scale through Liquid's sales team. Research, education, and non-profit use is always free with no revenue limit. There is no copyleft, so your fine-tunes stay yours. Every new organization also gets $2 in free API credit on the Liquid Console. And the weights themselves are on Hugging Face today, with native support for llama.cpp, MLX, and vLLM across Apple, AMD, NVIDIA, and Qualcomm hardware.

So the real price of LFM3 right now is the price of your own hardware and your own time. A short fine-tune is cheap in Liquid's own demos: turning LFM3-VL-2B into a wind-turbine-blade inspection model took about 3,000 public photos and a few hours on a single A100, and it lifted defect classification accuracy from under 10 percent out of the box to 67 percent, against GPT-6 Luna at under 10 percent on the same task. A robot-arm LoRA adapter took about six GPU-hours on a single AMD Instinct MI325X. If you were planning to call a hosted LFM3 API and pay per token, that option is announced but not priced — check the pricing page before you budget for it, because the rate does not exist yet.

Pros and cons

Pros: genuinely open weights under a clear, permissive license with a $10M revenue line; six specialized models on one backbone, so you download the modality you need instead of a generalist; agent and tool-use benchmarks that beat models five times its size; the fastest decoding and smallest KV cache in its size class on Liquid's measurements; decision models that answer in milliseconds with zero output tokens; day-0 support across llama.cpp, MLX, vLLM, Ollama, LM Studio, and Unsloth.

Cons: no published per-token API price for any LFM3 model, and no OpenRouter listing yet despite the launch-partner announcement; every benchmark is company-run with no independent head-to-head; knowledge and long-context scores trail bigger models by wide margins; English speech recognition is slightly worse than its own predecessor; audio tool-calling trails Gemma 4 E4B; the Decision Index is Liquid's own benchmark; and the $10M license line means a growing company eventually pays an unpublished enterprise price.

Who should use it, and who should skip it

Use LFM3 if you are building a specialized, on-device product: a support-triage pipeline, a document or GUI agent, a voice assistant that runs offline, a robot controller, or a fine-tuned expert for one narrow task. The license, the size, and the speed all point the same way, and the wind-turbine demo is the honest template — a small model, a few thousand examples, a few hours on one GPU, and a task it now does better than a frontier model that never specialized. It is the same playbook behind Recursive Flint, the free 40B laptop coding model we reviewed last week.

Skip it if you need general knowledge, long-document reasoning, or a hosted API with a published price. A 1.7B model will not out-know a 9B model, the 128K retrieval scores say so plainly, and the API rate card is empty. If your product is a general chatbot, a bigger hosted model is still the right call. And if you are a company approaching $10 million in revenue, get the enterprise license terms in writing before you build on the free tier, because the price above that line is not published.

FAQ

Is Liquid AI LFM3 free?
The weights are free to download, fine-tune, and use commercially under the LFM Open License v1.0 if your company earns under $10 million a year. Research, education, and non-profit use is always free. Above $10 million in annual revenue you need a paid enterprise license. The API has no published per-token price for LFM3 yet.

How many models are in the LFM3 family?
Six in the first release, all branched from the same LFM3-1.7B-Base backbone: LFM3-1.7B for text, LFM3-VL-2B for vision, LFM3-Audio-2B for audio, LFM3-Omni-2B for all three, and two decision models, d1-2B and d1-omni-2B.

Can LFM3 run on a laptop or a phone?
Yes, that is the design goal. Liquid tested across 14 device and backend combinations including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, an Apple M4 Max laptop, and AMD Ryzen AI laptops, with native support for llama.cpp, MLX, and vLLM.

What is a decision model, and how is d1-2B different from a chatbot?
A decision model does not write text. You give it a situation and a list of questions, and it returns calibrated, typed answers — yes/no, a choice, a score, a bounding box, an extracted span, or a number with a range — in one forward pass with zero output tokens. It is for routing, triage, moderation, and guardrails, not conversation.

Is LFM3 better than Qwen3.5 or Gemma 4?
On Liquid's own agent and tool-use benchmarks, the 1.7B text model beats both, and the vision model matches a 12B Gemma. On knowledge, long-context retrieval, and agentic coding, the bigger models win. There is no independent head-to-head yet, so treat the comparison as company-reported.

Key takeaways

— Liquid AI launched LFM3 on October 27, 2026: six open-weight models around 2B parameters on one shared backbone, covering text, vision, audio, omni, and decisions.
— The text model beats a 9B Qwen on Liquid's agent benchmarks but trails badly on knowledge and long-context tests, so it is an agent specialist, not a generalist.
— The vision model is the strongest card, matching a 12B Gemma across 62 benchmarks and leading on GUI understanding, grounding, and document work.
— The weights are free under $10M annual revenue with no copyleft, but no LFM3 API price is published and OpenRouter lists no LFM3 model yet.
— The decision models answer in milliseconds with zero output tokens, and the new typed answers — point, read, estimate, pick — make them usable for real triage pipelines.

As of October 28, 2026, LFM3 is the most interesting small-model launch of the week — it shares the week with Mistral Large 4, the 1T open model we reviewed on launch day — and also the least priced: download it free, fine-tune it on your own data, and run it on your own hardware — but do not budget for a hosted API until Liquid publishes a rate.