An unmerged SGLang pull request describes four Liquid AI LFM3 model variants spanning text, vision, audio and multimodal input. If the code reflects the family Liquid AI is preparing, it offers a concrete glimpse of what may be next, but it is not a launch announcement or proof of what will ship.
Watch Desk analysis
What happened
OrcaRouter reports that a pull request opened on 27 October adds SGLang support for LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B and LFM3-Omni-2B. The described models take combinations of text, images, video and audio as input, with text output. The audio variant is also wired for transcription.
The report says the pull request remains unmerged and its three CI checks are failing. It describes the checkpoints as developmental, and says Liquid AI has not announced LFM3. The listed sizes and capabilities are therefore details in a proposed software change, not confirmed product specifications.
Why it matters
The pull request sketches a compact multimodal family, including a variant intended to process video and its soundtrack. That is useful early evidence for developers tracking potential model support across inference software. It is also a reminder that infrastructure work can expose a project’s shape before a company is ready to present a product.
There are important blanks. The report says the code does not establish a release date, price, licence, benchmark results or total context length. It also does not show that the listed variants will be released as described. In particular, transcription support does not establish that the audio model can generate speech.
Our read
Interesting, yes; shopping list, no. An unmerged pull request is a worthwhile signal, especially when it names several variants and their input types, but it is a shaky foundation for product plans. Treat this as a reported preview of development work, not a promise from Liquid AI.
What to watch
- Whether the pull request is merged and its failing checks are resolved.
- Whether Liquid AI announces LFM3 or publishes model cards with confirmed specifications.
- Whether later documentation clarifies audio generation, context length, benchmarks and licensing.
Discussion spark: Should an unmerged code change be treated as useful early notice of an AI model family, or is it too provisional to guide anyone’s expectations?
Sources and evidence
- LFM3 Leak: Four Omni Models Liquid AI Hasn't Announced – OrcaRouter (27 October 2026, 15:31 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.