Liquid AI Launches LFM3 as Six Open-Weight Models for Text, Vision, Audio and Decisions

Liquid AI released six LFM3 models on October 27

Liquid AI announced LFM3 on October 27, 2026 and published public, ungated repositories for a six-model family: LiquidAI/LFM3-1.7B for text; LiquidAI/LFM3-VL-2B for text, images and video; LiquidAI/LFM3-Audio-2B for text and audio; LiquidAI/LFM3-Omni-2B for combined text, visual and audio work; and LiquidAI/d1-2B plus LiquidAI/d1-Omni-2B for structured, calibrated decisions. The base checkpoint is also available for post-training.

This is an artifact release, not a promise of weights later. The repositories expose model cards, configurations, Safetensors files and code, while Liquid AI supplies Transformers examples and fine-tuning notebooks. Availability still has layers: a downloadable checkpoint, a working local runtime, a Playground session and a priced production API are separate contracts.

One backbone branches into different modality contracts

The text model uses a 1.73B-parameter, 24-layer tri-hybrid backbone: 13 gated short-convolution layers, five sliding-window-attention layers and six global-attention layers. Liquid AI says it pre-trained the base on 22 trillion tokens and devoted the final 3.5 trillion-token stage to curated data with more agent traces and tool calls. Those training figures are provider disclosures, not an independently audited corpus account.

The vision model adds an in-house visual encoder; the audio model adds a causal audio encoder, decoder and LiquidCodec; the omni model combines both paths. Current repository metadata reports about 1.73B Safetensors parameters for text, 1.92B for vision-language, 2.00B for audio and 2.17B for omni. Exact capabilities still differ: the omni card says it is non-reasoning, while the text and vision cards document reasoning behavior.

Context and output limits differ across the family

LFM3-1.7B advertises 262,144 tokens across input and output. Its configuration confirms max_position_embeddings of 262,144. The model card positions it for local agents, tool calls and structured output, but also says the small model is not recommended for agentic coding or complex multi-step reasoning. A 256K capacity therefore should not be presented as 256K retrieval accuracy.

The reviewed LFM3-VL-2B and LFM3-Omni-2B cards specify 32K combined context. Omni accepts text, images, video with or without soundtrack and 16 kHz audio, then emits text, 24 kHz speech or interleaved speech and text. Media sampling, tokenization and output all consume the operational budget; teams should test their actual image sizes, frame rates, audio durations and output modes.

The benchmark and edge claims are test hypotheses

Liquid AI says LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters and reports 38.42 versus 3.02 on WebArena-Verified. It says LFM3-VL-2B raises its 62-benchmark average from 57.6 for the prior 3.1B model to 71.4, and reports its omni model ahead of selected models up to 10.7B parameters on three audio-visual benchmarks. The company also reports a 96.2% average across four LIBERO suites for a separately adapted robotics model.

These are vendor-run comparisons with different model classes, harnesses and post-training. AccessAllGPT did not inspect every prompt, trajectory, exclusion or scoring decision, and did not reproduce any result. Treat them as reasons to evaluate—not as proof of workload reliability, safety or superiority. Include failure cases, equal tool budgets, repeated runs and a small credible baseline.

Local deployment requires custom code and exact runtime checks

The model cards use trust_remote_code=True in Transformers examples. Text checkpoints also list routes for vLLM, SGLang, llama.cpp, Ollama, MLX and ONNX, while multimodal support varies by model and output mode. Pin the repository commit and dependency versions, inspect the custom code, and test the precise engine before assuming that one family member’s runtime support transfers to another.

Liquid AI says the text model fits in about 1 GB at 4-bit precision and every family member can be fine-tuned on a single GPU. Those are useful sizing claims but not a production bill. Measure model and processor bytes, KV cache, media encoder memory, startup, time to first token, streaming behavior, concurrency, energy and quality loss under the exact quantization and device.

The license places a $10 million boundary on commercial use

Every reviewed model card links LFM Open License v1.0 rather than Apache, MIT or another unrestricted permissive license. The agreement defines its threshold as annual revenue of $10 million or more and states that commercial use by a legal entity exceeding that threshold is not licensed under the agreement. The legal-entity definition includes entities under common control.

The license also carries redistribution, attribution and modification-notice conditions. “Open-weight” accurately describes artifact access; it does not mean every company can use the weights commercially without another agreement. Confirm the entity, revenue, use, derivative and distribution path with qualified counsel. This article is not legal advice.

API availability is announced, but public LFM3 pricing is missing

Liquid AI’s launch says the base and post-trained models are available through Hugging Face, its Playground and the Liquid AI API. However, the public pricing page reviewed October 28 listed d1, d1-ultra and LFM2.5-24B-A2B only. It did not publish an LFM3 API identifier, per-token price, cached-input rate or output rate.

Do not infer LFM3 pricing from the older models. Before building against managed access, verify the callable model ID in the account, supported inputs and outputs, context and file limits, rate limits, region, retention and training-use policy, service level and dated price. If those are unavailable, compare only the downloadable route or wait for a stable API contract.

What builders should do next

Trial one model, not “LFM3” as a bundle. Choose text when local tool use is central, VL when images or video are required, audio for speech-only flows, omni only when joint sight and sound removes real pipeline complexity, and d1 when calibrated structured decisions fit the task. Pin the model and runtime commits, review rights, scan custom code and establish a baseline before fine-tuning.

Constrain early tests to read-only or reversible work. Wait if the API contract, commercial rights, runtime support or device fit remains unresolved. Reject the model when representative quality, latency, memory, energy, safety or cost gates fail. A small model’s most valuable result may be proving that a narrow local workflow works without a frontier-scale service; that conclusion still needs local evidence.

Disclosures

AccessAllGPT received no weights, API or Playground access, credits, hardware, data, benchmark logs, license, briefing, review, payment or compensation for this article. Liquid AI and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank LFM3 or its hosting routes. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Liquid AI or Hugging Face. Publication-wide relationships are listed on the disclosures page.