Releases·Model Release
Liquid AI ships LFM2.5-VL-DSpark drafter, speeding LFM2.5-VL-3B decoding up to 3.13x on-device and 2.66x on H100
Liquid AI released LFM2.5-VL-DSpark, a 280M-parameter vision drafter that accelerates the LFM2.5-VL-3B model via speculative decoding, with day-one support in llama.cpp, MLX-VLM, and SGLang. It delivers decode speedups up to 3.13x on-device and 2.66x on H100, with end-to-end gains up to 2.62x and 2.27x respectively. The drafter is open-weight and available on Hugging Face in Safetensors and GGUF formats.
0 engagement1 sourceThu, Sep 24, 2026, 02:08 PM
AI Overview
- Liquid AI released LFM2.5-VL-DSpark, a speculative-decoding draft model adding 280M parameters (8.9%) to LFM2.5-VL-3B, with day-one llama.cpp, MLX-VLM, and SGLang support
- An attention-only 4-layer drafter with block size 9 runs image and text tokens through a shared hidden-state representation, reusing the text LFM2.5-DSpark algorithm unchanged
- Measured on six vision tasks: M5 Max decode speeds 2.30x-3.13x (end-to-end 1.56x-2.62x); M3 Ultra 1.57x-2.14x decode; H100 decode up to 2.66x and end-to-end up to 2.27x
- Speculation accelerates only decoding, not vision encoding or prefill, so end-to-end gains are capped by Amdahl's law on edge devices where prefill dominates latency
- The drafter is exact: the target verifies every proposed token, so greedy output matches the target model alone; weights ship on Hugging Face in Safetensors and GGUF formats
Related
Releases·Tool Release
Liquid AI merges DSpark support for LFM2 models into llama.cpp
Liquid AI's DSpark inference optimization for LFM2 models has been merged into llama.cpp via PR #27383, enabling up to 3.2x faster inference. Official GGUF files are available on Hugging Face, and the models are small enough to run on any hardware.
Liquid AI's DSpark inference optimization for LFM2 models has been merged into llama.cpp via PR #27383, enabling up to 3.2x faster inference. Official GGUF files are available on Hugging Face, and the models are small enough to run on any hardware.
31 engagement2 sourcesredditThu, Aug 20, 2026, 05:13 PM
Developments·Paper
SP$^2$ec: Training-Free Bandit Algorithm Adapts Self-Speculative Decoding for Vision-Language Models
An arXiv paper introduces SP²ec, a training-free bandit algorithm that adapts draft configuration during self-speculative decoding for vision-language models. It minimizes wall-time regret per query and improves throughput of existing self-speculative methods by up to 11%.
An arXiv paper introduces SP²ec, a training-free bandit algorithm that adapts draft configuration during self-speculative decoding for vision-language models. It minimizes wall-time regret per query and improves throughput of existing self-speculative methods by up to 11%.
0 engagement1 sourcearxivTue, Oct 13, 2026, 05:08 PM
Releases·Model Release
LiquidAI releases LFM2.5-Encoder-350M, a 15-language bidirectional encoder built on LFM2 for on-device fine-tuning
Liquid AI released LFM2.5-Encoder models in 230M/250M and 350M sizes, bidirectional masked language models built on the LFM2 architecture. They target fine-tuning into task-specific models for classification, retrieval, reranking, and semantic similarity across 15 languages, with GGUF releases and a llama.cpp PR enabling browser and CPU use.
Liquid AI released LFM2.5-Encoder models in 230M/250M and 350M sizes, bidirectional masked language models built on the LFM2 architecture. They target fine-tuning into task-specific models for classification, retrieval, reranking, and semantic similarity across 15 languages, with GGUF releases and a llama.cpp PR enabling browser and CPU use.
5 engagement1 sourceredditSat, Oct 3, 2026, 07:00 AM
Releases·Model Release
LiquidAI ships d1-3B and d1-omni-600M decision models with zero output tokens
Liquid AI released Open d1, two open-weight multimodal decision models: d1-3B and d1-omni-600M. Both answer a set of typed questions about a given state in a single forward pass with zero output tokens, targeting low-latency uses like agent guardrails, reranking, and visual inspection. d1-3B tops the Decision Index v0.2.1 among models under 10B parameters.
Liquid AI released Open d1, two open-weight multimodal decision models: d1-3B and d1-omni-600M. Both answer a set of typed questions about a given state in a single forward pass with zero output tokens, targeting low-latency uses like agent guardrails, reranking, and visual inspection. d1-3B tops the Decision Index v0.2.1 among models under 10B parameters.
11 engagement2 sourcesredditWed, Oct 7, 2026, 05:02 PM
Developments·Paper
H-Spec: Hybrid Mamba-Attention Drafter Cuts Drafter-Side KV Cache for Parallel Speculative Decoding
An arXiv paper introduces H-Spec, a hybrid Mamba-attention parallel drafter for speculative decoding that eliminates the separate drafter-side KV cache. It reuses target KVs in place and injects target hidden states only at the last input position, improving accepted length and latency across three target models. Under concurrent serving it delivers higher throughput with lower KV cache utilization than baselines.
An arXiv paper introduces H-Spec, a hybrid Mamba-attention parallel drafter for speculative decoding that eliminates the separate drafter-side KV cache. It reuses target KVs in place and injects target hidden states only at the last input position, improving accepted length and latency across three target models. Under concurrent serving it delivers higher throughput with lower KV cache utilization than baselines.
0 engagement1 sourcearxivMon, Sep 21, 2026, 07:11 AM