Open Source · OCTOBER 27, 2026
Liquid AI ships LFM3: six 2B-class open-weight models, fine-tunable on one GPU
The October 27 release spans text, vision, audio, omni, and two decision models branched from a 1.7B tri-hybrid base, with a ~1 GB 4-bit footprint and a hosted text API listed at $0.08 per million input tokens.
Liquid AI released LFM3 on October 27, 2026: six open-weight models in the 2B class, all branched from a 1.73B tri-hybrid backbone, with a hosted text API listed at $0.08 per million input tokens and weights small enough to fit in roughly 1 GB at 4-bit precision. The move pushes the frontier of what "specialized intelligence" means at the small-model tier, and it does so by collapsing the deployment floor rather than chasing capability headlines.
The family spans text (LFM3-1.7B, 16 languages, 128K vocabulary), vision (LFM3-VL-2B, 1.9B parameters), audio (LFM3-Audio-2B, 18 languages), an omni variant (LFM3-Omni-2B, 2.2B parameters), and two decision models. All share the same 24-layer backbone: 13 gated short-convolution layers with a kernel size of 7 (up from 3 in LFM2), 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Training moved 22 trillion tokens through pre-training, 3.5 trillion through mid-training (which stretched the context to 256K), 1.75 trillion through supervised fine-tuning, and a four-stage post-training regime that leaned on six specialist RL teachers for on-policy distillation.
The economic story is in the KV cache. At 128K tokens, LFM3-1.7B's cache lands at 1.6 GB. Liquid's own figures put Qwen3.5-4B at 4.3 GB and MiniCPM5-2B at 5.6 GB in the same slot. That's the number that decides whether an inference server fits on commodity hardware or a Raspberry Pi.
Headline benchmarks: WebArena-Verified 38.42 (versus 28.94 for Qwen3.5-9B and 3.02 for Liquid's own LFM2.5-2.6B), BFCL V4 function calling at 71.39, AutomationBench public at 41.29, IFEval at 96.66, GPQA Diamond at 63.37. Losses are honest: MCP-Atlas 32.57 against a competitor's 42.67, SWE-bench Verified 43.40 against 51.20. Every row is Liquid's harness, which is the asterisk the entire release carries until Artificial Analysis replicates.
Pricing scales with modality. LFM3-VL-2B runs $0.10 in and $0.50 out per million tokens. Audio and omni are $0.10 in and $1.00 out, with separate audio-token charges of $2.00 input and $10.00 output per million. A full million-in, million-out run on the text model costs $0.28. The LFM Open License v1.0 keeps commercial use royalty-free below $10 million in annual revenue.
Context, finally, is where the release gets quietly interesting. The hosted API caps at 128,000 tokens; the downloadable weights go to 256K. For founders already evaluating full-duplex small-footprint speech stacks, LFM3 is the text and vision companion that lives on the same box.
Sources
- https://www.liquid.ai/blog/lfm3
- https://venturebeat.com/technology/liquid-ais-new-lfm3-models-put-browser-agents-voice-commands-and-image-recognition-onto-enterprise-controlled-devices
- https://www.unite.ai/liquid-ai-launches-lfm3-family-of-six-open-weight-2b-class-models/
- https://aiweekly.co/alerts/liquid-ai-opens-six-model-lfm3-family-at-256k-context
- https://www.orcarouter.ai/blog/lfm3-family-release-tri-hybrid