Models & Labs
llama.cpp b12342 adds few-row MMQ for speculative decoding
llama.cpp Releaseshigh confidence
Why it matters
- →Optimizes speculative decoding verification passes for faster local inference.
- →Adds support for quantized formats on NVIDIA GPUs from Ampere onwards.
- →Reduces latency in token generation without requiring model changes.
llama.cpp has released build b12342, featuring a new CUDA kernel optimized for few-row matrix multiplications. This update specifically targets the verification passes in speculative decoding, where source tensors have between 2 and 24 columns. The implementation supports multiple quantization formats including Q8_0, Q3_K, Q4_K, Q5_K, Q6_K, IQ3_S, and IQ4 variants on NVIDIA GPUs starting from the Ampere architecture. Alongside this performance enhancement, the release includes minor fixes for weight flags in testing and clearer error handling for HIP/MUSA backends.
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
- 1llama.cpp adds OpenCL iq3 quantization support
llama.cpp Releases · November 1, 2026 · Same story
- 2llama.cpp b12122 fuses NVFP4 matmuls for CUDA
llama.cpp Releases · November 5, 2026 · Same story
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12336 fixes HunyuanVL crash
This release quietly patches a critical memory overflow in the mtmd module that caused crashes when loading HunyuanVL models. The issue stemmed from F64 position embeddings overflowing an F32 buffer, a subtle bug that broke multimodal inference for users relying on this specific architecture. By enforcing strict F32 requirements, llama.cpp stabilizes support for Tencent's vision-language model without adding new features or changing the core API. It’s a necessary maintenance update that keeps local multimodal workflows from breaking unexpectedly.
llama.cpp Releases
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
Coding Toolscoding
llama.cpp b12343 adds qwen3.5 and ROCm 10
This release quietly extends llama.cpp’s hardware coverage with ROCm 10.0 binaries for both Linux and Windows, finally bringing AMD RDNA/CDNA users into the fold without manual compilation. The standout feature is native support for Qwen3.5 models via tensor merging during conversion, a necessary step as this new architecture gains traction in the open-source community. By shipping pre-built CUDA 13 binaries alongside the older 12.x versions, the project ensures forward compatibility with newer NVIDIA GPUs out of the box. There are no new quantization algorithms or performance breakthroughs here; it is purely an infrastructure update to keep pace with the rapidly shifting landscape of model architectures and GPU drivers.
llama.cpp Releases
More in Models & Labs
© 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
© 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
© The AI Daily BriefModels & Labsmodels
Google Releases Gemini 4 Carbon Model
Google has released the Gemini 4 Carbon model, expanding its multimodal AI lineup with new capabilities for developers and users.
The AI Daily Brief