Liquid AI released a 279.5M-parameter attention-only drafter for LFM2.5-VL-3B. The target model verifies every proposal, so greedy output remains unchanged. Across six MMSpec vision tasks, Liquid reports decode speedups of 2.30-3.13x on M5 Max, 1.57-2.14x on M3 Ultra and up to 2.66x on H100; end-to-end gains are smaller because image encoding and prefill do not accelerate. Safetensors and GGUF weights are available with llama.cpp, MLX-VLM and SGLang integrations. The release was recovered from overlap.
Liquid AI releases a speculative drafter for vision-language decoding
2026-09-24Verified announcementTRIAL
Liquid AI released a 279.5M-parameter attention-only drafter for LFM2.5-VL-3B. The target model verifies every proposal, so greedy output remains unchanged. Across six MMSpec vision tasks, Liquid reports decode speedups of 2.30-3.13x on M5 Max, 1.57-2.14x on M3 Ultra and up to 2.66x on H100; end-to-end gains are smaller because image encoding and prefill do not accelerate. Safetensors and GGUF weights are available with llama.cpp, MLX-VLM and SGLang integrations. The release was recovered from overlap.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams already serving LFM2.5-VL, this is a low-risk latency experiment because the target verifies every emitted token. Benchmark full request latency on your image sizes and output lengths; decode-only gains will overstate benefits for short answers or heavy vision prefixes.
Evidence & related material
Technical details & scores
| Drafter | 279.5M parameters; 8.9% of the 3B target; attention-only |
|---|---|
| Block Size | 8-9 |
| Decode Speedup | 2.30-3.13x M5 Max; 1.57-2.14x M3 Ultra; up to 2.66x H100 |
| End To End | up to 2.62x M5 Max; 1.77x M3 Ultra; 2.27x H100 |
| Runtime | llama.cpp, MLX-VLM and SGLang |
| Limit | vision encoding and prefill are unchanged |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 2
Verification cost 3
Risk 2