Coding Tools
llama.cpp b12059 optimizes unified KV cache for multi-sequence
llama.cpp Releaseshigh confidence
Why it matters
- →Sparse KV cache handling reduces latency for multi-turn agent workflows.
- →Unified batch processing avoids performance penalties from interleaved sequences.
- →Maintains graph topology efficiency without falling back to dense attention masks.
llama.cpp has released version b12059, focusing on optimizing the unified key-value (KV) cache for multi-sequence attention. The update replaces dense KQ masks with sparse row lists built from per-sequence cell bitmaps, improving efficiency for sliding window attention and interleaved batches. This change preserves graph topology for non-equal-length runs, reducing computational overhead in complex inference scenarios. The release includes binaries for macOS, Linux, Windows, and Android across CPU, CUDA, ROCm, and Vulkan 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 b10662 release refines KV cache handling
llama.cpp Releases · August 28, 2026 · Same story
- 2llama.cpp b12065 fixes KV cache graph scheduling crash
llama.cpp Releases · November 3, 2026 · Same story
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12053 updates Apple memory and adds ROCm 10
This release quietly fixes a memory allocation quirk on Apple Silicon by switching to vm_allocate, which should stabilize buffer handling across macOS and iOS. More notably, it brings ROCm 10.0 support to both Linux and Windows, finally giving AMD GPU users parity with CUDA builds in the standard distribution. The update also extends CUDA 13 compatibility to ARM64 architectures, broadening the hardware base for high-performance inference. It’s a maintenance-heavy release that solidifies llama.cpp as a cross-platform runtime rather than just a local Mac tool.
llama.cpp Releases
Models & Labsother
llama.cpp b12054 adds CUDA 13 and ROCm 10.0
This release quietly cements llama.cpp as the universal inference runtime by adding default support for CUDA 13 and ROCm 10.0. NVIDIA users can now leverage newer driver stacks without waiting for separate builds, while AMD GPU owners finally get parity with modern ROCm versions. The inclusion of Snapdragon and OpenVINO binaries further widens the hardware net, though Apple Silicon KleidiAI is notably disabled here. It’s less about new features and more about ensuring no major accelerator is left behind in the local LLM ecosystem.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12057 adds ROCm 10 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally bringing first-class ROCm 10.0 support to both Linux and Windows. AMD GPU users no longer need workarounds, effectively closing a long-standing parity gap with NVIDIA's CUDA ecosystem. Equally notable is the new Snapdragon build for Linux, which unlocks local AI on ARM-based laptops using Adreno GPUs and Hexagon NPUs. While KleidiAI on Apple Silicon is temporarily disabled, the expansion into AMD and Qualcomm hardware signals a strategic shift toward broadening accessible local inference beyond just NVIDIA.
llama.cpp Releases
More in Coding Tools
Coding Toolscoding
Claude Code v2.1.317 patches security and stability
This release prioritizes trust and reliability over new features, addressing critical friction points for enterprise users. The most significant addition is a safety check that prompts before sending account emails to external networks, directly tackling data leakage concerns in automated workflows. For developers managing scale, the new gateway admin API cost report provides granular visibility into spend per developer and IdP group, solving a major blind spot in self-hosted deployments. Stability fixes for session persistence, proxy timeouts, and credential handling ensure that long-running agent tasks don't silently fail or leak data.
Claude Code Releases
© Hugging Face BlogCoding Toolscoding
Hugging Face introduces Model Integration Contracts
Hugging Face is solving the regression problem in its massive Transformers library by introducing Model Integration Contracts (MICs). Instead of relying on slow full-scale GPU tests, they shrink real checkpoints into tiny CPU-friendly fixtures that preserve architectural quirks. This approach caught a two-month-old bug in T5’s attention mechanism that standard forward passes missed entirely. It shifts testing from brute-force coverage to precise, reproducible contract verification for every model architecture.
Hugging Face Blog
© The Verge AICoding Toolscoding
Anthropic blocks Claude Code from cloning existing games
Anthropic has tightened Claude Code’s system prompt to explicitly forbid recreating specific existing video games, targeting the surge in AI-generated decompilations and ports. The update on Opus 5.5 draws a hard line against IP infringement while preserving the ability to build original titles or genre-inspired projects from scratch. This marks a significant shift in how coding assistants handle intellectual property boundaries, moving from passive compliance to active prevention of direct clones. Developers can now expect stricter refusals when attempting to replicate protected game mechanics or assets.
The Verge AI