Hugging Facexx, Yuri Khrustalev, Leonie Monigatti, Viviana MárquezAnnouncement
Liquid AI released an experimental DSpark draft model for their vision-language model LFM2.5-VL-3B. The drafter adds a speculative decoding path that trades a minimal increase in memory footprint for a larger speedup without changing output quality. The 280M-parameter drafter, a 4-layer attention-only transformer, captures the target model’s hidden states at fixed tapped layers to draft a block of candidate tokens. On an M5 Max, decoding runs 2.30x to 3.13x faster by task, with end-to-end latency improvements of 1.56x to 2.62x. The drafter ships with day-one support for llama.cpp, MLX-VLM, and SGLang, and is available on Hugging Face in Safetensors and GGUF formats.
Liquid AI released an experimental DSpark draft model for their vision-language model LFM2.5-VL-3B. The drafter adds a speculative decoding path that trades a minimal increase in memory footprint for a larger speedup without changing output quality. The drafter has approximately 280M parameters and increases the deployed model’s parameter count by just 8.9%.
The DSpark draft model is a 4-layer attention-only transformer with 280M parameters that captures the target model’s hidden states at fixed tapped layers and conditions on them to draft a block of candidate tokens. Image patches and text tokens are projected into a shared representation before those layers, so the drafter operates on hidden-state vectors of identical dimensionality regardless of input modality. The inference algorithm is unchanged from the text models.
The DSpark draft model for LFM2.5-VL-3B adds 280M parameters (8.9% increase) and provides decode speedups up to 3.13x on device and 2.66x on an H100.
Speculative decoding speeds up only the decode phase, not vision encoding or prefill, so end-to-end gains are capped by Amdahl’s law when those stages dominate wall time.
The vision drafter uses the same architecture as text drafters, projecting image patches and text tokens into a shared representation so the drafter operates on hidden-state vectors of identical dimensionality.
Speculative decoding speeds up only the decode phase, not vision encoding or prefill. When those stages already take up much of the wall time, even a large decode speedup gives only a modest end-to-end gain. This is Amdahl's law, where the overall speedup is capped by the part of the workload that isn't accelerated.
The drafter adds 280M parameters, which is 8.9% on top of the 3B target model. The decoder stack (4 layers) is 193.0M, hidden-state projection is 21.0M, Markov head is 65.5M, and norms plus confidence head is 6.4k.
The vision drafter uses the same architecture as the text LFM2.5-DSpark drafters. Image patches and text tokens are projected into a shared representation before the tapped layers, so the drafter operates on hidden-state vectors of identical dimensionality regardless of input modality.
Summaries are generated from the original by a language model. Check the original before relying on it.
See the idea
Blackwell inference is a pipeline, not a chip upgrade. This explainer walks five stages — fleet validation, batching, KV caching, low-precision kernels, and speculative decoding — that turn peak FLOPs into cheap tokens.
bfcache freezes page snapshots for instant back/forward navigation, but unload listeners, cache headers, or third-party scripts can silently disable it. This explainer maps the trade-offs and the metrics that reveal them.
Keep reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
One useful email. Unsubscribe anytime.
