One signal

Update What changed

Liquid AI’s expanded LFM3 announcement says its text model fits in about 1 GB at 4-bit precision and has a 256K-token context window. The company says LFM3-1.7B is designed for local agents and tool use, and can be fine-tuned on a single GPU.

The post also gives comparison results: Liquid AI says LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters. At 8K tokens of context, it reports the model is about 25% faster than LFM2-1.2B. These are company-reported results, not independent benchmarks.

For multimodal use, Liquid AI says LFM3-VL-2B accepts video, while its audio model supports speech recognition, synthesis and tool calls. The company also describes LFM3-Omni-2B as handling text, images and audio in one model, with both text and speech output.

Sources and evidence

• Build specialized multimodal models with LFM3: Liquid AI’s expanded LFM3 post describes local deployment and fine-tuning options, model capabilities, and comparative performance claims.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Liquid AI’s expanded LFM3 notes add benchmark figures beyond the model’s launch announcement. The company says LFM3-1.7B scored 38.42 on WebArena-Verified, compared with 3.02 in its stated comparison, and says it beats its LFM2.5-2.6B predecessor while using 35% fewer parameters.

The company also reports that its vision model averages 71.4 across 62 vision benchmarks, up from 57.6 for LFM2.5-VL-3B. LFM3-VL-2B is its first vision model to accept video, while the audio model’s first-token latency can be as low as 50 milliseconds, according to Liquid AI.

For people looking to try the models, Liquid AI says they run in Transformers; its guidance gives a basic setup for LFM3-1.7B and points to model cards for further inference options. These are company-reported capabilities and benchmark results, not independent evaluations.

Sources and evidence

• Liquid AI | Device-native foundation models.: Liquid AI’s expanded LFM3 announcement supplies additional model-specific benchmark results, reported performance details and guidance for running the models.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

MarkTechPost says LFM3’s free commercial use is limited by a $10 million annual revenue threshold. That is a significant qualification for businesses weighing the models: “open-weight” does not necessarily mean free commercial use at every scale.

The report says Liquid AI supports GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints, with runtimes including llama cpp, MLX, vLLM, SGLang, Ollama and Transformers. It also says the models were tuned across 14 device and backend combinations, including a Raspberry Pi 5 and Qualcomm’s Hexagon NPU.

For local use, MarkTechPost says LFM3-1.7B has 1.73 billion parameters and a 256K context window; it reports that a full 128K context needs 1.6GB of KV cache. Those details make the deployment trade-offs more concrete, while the performance figures in the report remain claims attributed to Liquid AI rather than independent evaluations.

Sources and evidence

• Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models - MarkTechPost: MarkTechPost says Liquid AI’s LFM3 models support free commercial use only up to a $10 million annual revenue threshold and reports additional deployment formats, runtimes and model specifications.

Independent WittyWires Watcher; not an official account or feed.