Liquid AI releases LFM3 models, including a 1.7B model with 256K context

• Liquid AI released six edge models in the LFM3 family, according to maximelabonne's announcement.
• The 1.7B text model supports 256K context and takes about 1 GB at 4-bit precision, according to Ollama's announcement.
• Browser agents are the standout: LFM3 scores 38.4 on WebArena-Verified versus Qwen3.5-9B's 28.9 in Liquid's benchmark chart.
• An M4 Max generated 280 tokens per second in Magnitude, according to tomgreenwald's local test.

Liquid's training recipe includes an RL length penalty and six specialist teachers distilled back into one model. Its vision encoder packs pairs of sampled video frames into a single temporal slice.

Liquid's October 27 release branches from LFM3-1.7B-Base:

• LFM3-1.7B: text input and output, post-trained for agent harnesses including Hermes Agent, Pi, and OpenClaw.
• LFM3-VL-2B: text, image, and video input; text output.
• LFM3-Audio-2B: text and audio input; text and speech output, with speech capabilities across 18 languages.
• LFM3-Omni-2B: text, image, video, and audio input; text and speech output.
• d1-2B: decisions from text, images, and video.
• d1-omni-2B: decisions with audio input added.

Both d1 variants answer multiple questions in one forward pass with calibrated probabilities. Liquid says every model can be fine-tuned on a single GPU.

Liquid's backbone description identifies four architectural changes:

1. Layer mix: 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six global attention layers. Convolution kernels grow from three to seven.
2. Head-wise gating: a sigmoid gate scales each attention head's output, helping suppress activation outliers that hurt quantization.
3. Partial RoPE: global attention rotates only the highest-frequency half of each head; sliding-window layers retain full RoPE.
4. Sandwich normalization: each block normalizes both its input and output.

Pre-training used 22 trillion tokens. The final 3.5 trillion emphasized curated data, agentic traces, and tool calls while extending context to 256K.

Liquid's evaluation notes say the comparison uses BF16 models served through vLLM. The approximately 1 GB deployment claim refers to 4-bit precision.

The broader table reports RULER 128K scores of 65.56 for LFM3 versus 87.11 for Qwen3.5-9B. WebArena-Verified allowed up to 40 steps; SWE-bench used OpenCode with up to 250 turns.

Liquid explicitly says the text model is not recommended for agentic coding or complex multi-step reasoning because of its small size.

The text model's 256K limit counts input and output together. Ollama's Audio specification and Omni specification instead list 32K in their readmes, although the Omni tag table separately labels context as 125K.

Omni is also explicitly listed as a non-reasoning model without thinking support.

Text, vision, audio, and omni checkpoints arrived on Ollama, which announced local availability at launch. Its text-model listing provides the command:

Magnitude also shipped day-one support:

In early use, anderslie reported fewer thinking loops than with small Qwen models, while placing LFM3 in a lower capability tier than larger models.

Ollama's vision-model page still marks video input as “soon,” although Liquid's launch inventory lists video support.

The text-model chat template supports tool discovery without putting every function schema into the initial prompt:

1. Tools marked defer_loading=True stay out of the rendered prompt.
2. The harness implements a search tool that returns names as tool_reference parts.
3. The template inserts those tools' definitions into <additional_tools>.

The complete tool list still accompanies every request. Deferred loading controls which definitions reach the model's prompt, while the application implements and executes the search.