Models & Labs
Liquid AI releases DSpark drafter for vision models
Hugging Face Bloghigh confidence
Why it matters
- →Speculative decoding drafter for vision-language models reduces latency significantly on edge hardware.
- →Day-one support for major inference engines (llama.cpp, MLX, SGLang) lowers integration friction.
- →Minimal parameter overhead (280M) makes it practical to deploy alongside existing VLMs.
Liquid AI has released LFM2.5-VL-DSpark, a speculative decoding drafter designed to accelerate vision-language models on both edge devices and datacenter GPUs. The open-weight drafter adds just 280M parameters (8.9% overhead) to the target model while achieving up to 3.13x faster decoding on Apple Silicon M5 Max chips and 2.66x speedups on NVIDIA H100s. It ships with immediate integration support for llama.cpp, MLX-VLM, and SGLang, allowing developers to deploy faster VLM inference without rewriting their inference stacks. The release underscores a growing industry shift toward optimizing inference efficiency through specialized drafters rather than solely increasing model scale.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Together AI Hosts MiniMax M3 for Efficient Inference
Together AI Blog · June 2, 2026 · Related
- 2vLLM v0.25.0 Release Enhances Model Execution
vLLM Releases · July 13, 2026 · Related
- 3vLLM v0.26.0 Release Enhances Model Support
vLLM Releases · July 27, 2026 · Related
- 4Startups Innovate Beyond Transformers in LLMs
MIT Technology Review AI · August 10, 2026 · Related
- 5LFM2.5-DSpark Boosts Inference Speed by Up to 3.2x
Hugging Face Blog · August 20, 2026 · Same story
- 6GLM 5.3 Flash: Z.AI's Cost-Efficient Model Variant
Sam Witteveen · August 30, 2026 · Related
- 7llama.cpp adds vision target support for DFlash/Dspark
llama.cpp Releases · September 24, 2026 · Related
What happened next
- 8Liquid AI releases LFM3 multimodal models
Hugging Face Blog · October 27, 2026 · Same story
- 9llama.cpp adds LiquidAI LFM3 model support
llama.cpp Releases · October 28, 2026 · Same story
- 10Together AI automates inference optimization with agents
Together AI Blog · November 4, 2026 · Related
More from Hugging Face Blog
© Hugging Face BlogModels & Labscoding
IBM releases LLMON markup to fix prompt injection
IBM is tackling the brittleness of small language models by introducing LLMON, a structured markup language that explicitly separates instructions from data. By using reserved special tokens instead of ambiguous characters like JSON brackets, these new Granite 4.1 preview models drastically reduce indirect prompt injection attacks while maintaining standard performance. This approach forces the model to bind specific constraints to specific data payloads, eliminating the confusion that often leads to hallucinations or security breaches in long contexts.
Hugging Face Blog
© Hugging Face BlogModels & Labsother
Hugging Face releases TRELLIS.2 fine-tuning guide
Hugging Face finally provides a concrete path to adapt the powerful TRELLIS.2 model for specific domains like footwear or furniture, solving the 'generalist' problem in 3D generation. By targeting only the structure and shape DiTs with LoRA adapters while freezing the texture flow, they prevent catastrophic forgetting of material priors—a common pitfall when fine-tuning large generative models. This approach allows practitioners to achieve high geometric accuracy on small datasets (188 samples) without retraining the entire pipeline. It transforms TRELLIS.2 from a demo toy into a usable tool for production-grade 3D asset creation.
Hugging Face Blog
© Hugging Face BlogOpen Sourceagents
Tether AI launches BotPsy, a sub-billion agentic model
Tether AI Research has released BotPsy, an 0.8B-parameter model built on Qwen3.5 that specializes in tool-use and agentic workflows for on-device deployment. It achieves state-of-the-art performance in its size class, outperforming larger models like Qwen3.5-2B-Think on benchmarks such as BFCL v4 and ACEBench while maintaining strict privacy by keeping data local. The model is designed to handle routine tasks like calendar management and web search without relying on expensive cloud APIs, offering a viable path for private, self-hosted personal assistants. With open weights under Apache 2.0 and GGUF support, it provides immediate utility for developers building lightweight, autonomous agents.
Hugging Face Blog
More in Models & Labs
Models & Labsother
llama.cpp b12340 adds CUDA 13 and ROCm 10.0
This release quietly closes the hardware gap for local inference by adding default support for CUDA 13 and ROCm 10.0 across Linux and Windows. Users with newer NVIDIA GPUs or AMD accelerators no longer need to compile from source to get performance, while Apple Silicon builds now ship KleidiAI disabled by default, shifting the focus back to standard ARM optimizations. The inclusion of Snapdragon support for Linux and Android marks a significant step toward on-device AI beyond the x86/ARM desktop ecosystem. It is less about new features and more about ensuring llama.cpp remains the universal runtime for every available accelerator.
llama.cpp Releases
Models & Labscoding
llama.cpp b12342 adds few-row MMQ for speculative decoding
Speculative decoding just got a significant speed boost on NVIDIA GPUs. This release introduces a specialized kernel for 'few-row' matrix multiplications, targeting the small batch sizes typical of verification passes in speculative decoding workflows. By optimizing these specific operations for quantized formats like Q8_0 and IQ4 variants, llama.cpp reduces latency where it matters most during token generation. The change is CUDA-only for now but leverages hardware capabilities from Ampere onwards, making local inference faster without requiring new model architectures or complex configuration changes.
llama.cpp Releases
© TechCrunch AIModels & Labsmodels
Google deploys Gemini 4 Hydrogen for faster Search
Google is quietly shifting its search infrastructure by replacing older fast models with Gemini 4 Hydrogen, a tier designed specifically for speed and cost-efficiency rather than raw reasoning power. This model completes tasks 2.5 times faster than its predecessor while beating the more powerful Argon on specific intelligence benchmarks, proving that specialized efficiency can outperform brute force in high-volume workflows. By integrating this into Search's fan-out technique, Google now breaks down complex queries into subqueries with greater precision and speed, delivering richer visual responses without the latency of heavier models. This marks a strategic pivot toward agentic execution layers within consumer search, making AI Mode feel snappier for everyday multi-part questions.
TechCrunch AI