LFM2.5-VL-DSpark
Liquid AI released open-weight DSpark draft models that bring exact speculative decoding to the LFM2.5-VL-3B vision model, with day-one support in llama.cpp, MLX, and SGLang.
Why it matters
Speculative decoding is exact here - the target verifies every proposed token - so it speeds up on-device vision-language inference with zero quality change; a concrete step toward real-time visual agents on consumer hardware.
What you could build with it
Build an on-device visual pantry-inventory agent on a Mac: LFM2.5-VL-3B with DSpark speculative decoding watches a phone camera feed and updates a shopping list in real time with no cloud calls.
Does it hold up?
Promising but unverified on real devices. Immediate third-party integration and exact greedy decoding are confirmed, but there are no real phone benchmarks yet and the $10M-revenue commercial-license threshold matters for production use.
Built with LFM2.5-VL-DSpark
- lablup/mlxcel integrationintegration · Third-party integration wiring the Liquid AI model family into an ML serving stack.
- Community models doc referencing the releasewriteup · Community documentation page tracking the Liquid AI model lineup including the DSpark variant.
- Licensing and performance scrutinywriteup · Independent scrutiny of the on-device speedup claims and the commercial license terms, including the $10M revenue threshold.
- lablup/mlxcelgithub · 470 ★ · Third-party ML serving integration for the model; stars not observed.
Learn more
First spotted on hf: source.
More AI models
ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69Mistral Forge: enterprise platform for training and continuously improving proprietary modelsMistral AI launched Forge (Sep 28) — a full-lifecycle model training platform (pre-training, SFT, DPO/ODPO, RL…models · JEV 0.69kev: open, trainable Jev-like family of small decision models on Qwen3.5/3.8A family of small decision models (0.8B to 27B) built on Qwen3.5/3.8 that reproduces Jev's prefill-only typed-decision…models · JEV 0.68GPT-Synopsys: OpenAI and Synopsys multi-year deal to build an AI model that operates EDA toolsOpenAI and Synopsys signed a multi-year agreement on Sep 30 to jointly develop GPT-Synopsys, a specialized model that…models · JEV 0.66
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.