Coding Tools
llama.cpp adds Hexagon IM2COL_3D support
llama.cpp Releaseshigh confidence
Why it matters
- →Optimizes 3D convolutions on Snapdragon NPUs, improving VLM performance on mobile devices.
- →Removes inefficient scalar FP16 conversions in the Hexagon kernel for better throughput.
- →Updates CUDA to 13.4 and adds ROCm 10.0 support for broader GPU compatibility.
llama.cpp has released an update featuring optimized Hexagon IM2COL_3D support for Qualcomm Snapdragon NPUs. The changes involve rewriting the kernel to utilize a DMA double-buffer pipeline, eliminating inefficient scalar FP16 conversions and divisors. This optimization aims to improve performance for 3D convolution operations common in vision-language models. The release also updates CUDA support to version 13.4 and adds ROCm 10.0 compatibility.
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 b11044 updates Hexagon and CUDA builds
llama.cpp Releases · September 19, 2026 · Same story
- 2llama.cpp b11054 adds Hexagon TOP_K support
llama.cpp Releases · September 20, 2026 · Same story
Follow this story
Open the full story →llama.cpp b11272 optimizes Hexagon NPU and adds CUDA 13
2 developments
- Sep 30 · llama.cpp Releases
llama.cpp b11272 optimizes Hexagon NPU and adds CUDA 13 - Oct 12 · llama.cpp Releases
llama.cpp adds Hexagon IM2COL_3D support (This article)↳ llama.cpp rewrites Hexagon IM2COL kernel to use canonical DMA double-buffer pipeline removing scalar FP16 conversions
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b11555 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. The inclusion of Snapdragon AI stack binaries for Linux marks a significant expansion into mobile-edge hardware, while the addition of CUDA 13 builds future-proofs the tool against upcoming driver requirements. By standardizing these complex backend integrations into single downloadable binaries, the project removes friction for anyone deploying local models on non-NVIDIA silicon.
llama.cpp Releases
Models & Labsmodels
llama.cpp adds Prism Bonsai 2 27B support
Prism's Bonsai 2 27B model is now natively supported in llama.cpp, bringing a high-efficiency architecture to the local inference ecosystem. The release includes specific runtime optimizations for its Hadamard transformations, ensuring smooth execution across diverse hardware backends like CUDA and ROCm. This integration allows developers to run a specialized 27B parameter model without custom conversion scripts or external wrappers. It expands the practical toolkit for those needing compact yet capable models on consumer-grade GPUs.
llama.cpp Releases
Coding Toolscoding
llama.cpp RPC logging overhaul
This release refactors llama.cpp’s RPC backend logging, turning a binary debug flag into a granular verbosity system. Developers can now trace tensor transfers and graph cache evictions with precision, which is critical for debugging distributed inference setups. The cleanup also unifies log handling across client and server sides, reducing noise and fixing format specifier bugs. It’s a maintenance win that makes diagnosing remote inference failures significantly easier.
llama.cpp Releases
More in Coding Tools
© GitHub ChangelogCoding Toolscoding
Copilot JetBrains update adds model controls and MCP settings
GitHub Copilot for JetBrains is tightening enterprise grip with managed default models, letting admins lock in specific agents while preserving user choice. The diagnostic workflow gets a direct Fix action that bridges inline chat and agent mode, cutting steps between error detection and resolution. Crucially, new MCP controls let users disable automatic server startup, addressing resource concerns around tool-heavy configurations. This release also drops support for IDE 2025.1, forcing an upgrade to 2025.2 or later.
GitHub Changelog
Coding Toolscoding
Claude Code v2.1.292 patches security and agent logic
This release tightens the leash on Claude Code's autonomous capabilities while fixing critical sandbox escapes. The new effort parameter for Agent tools lets developers explicitly control sub-agent depth, a necessary guardrail as these systems grow more complex. Security fixes are prominent, addressing how plugins handle network paths and how file permissions persist during session resumption. It’s a stability patch that ensures the tool remains usable in enterprise environments without compromising on the new agent features.
Claude Code Releases
Coding Toolscoding
Claude Code v2.1.293 updates Haiku and fixes agent bugs
Anthropic quietly shipped a significant model update alongside routine maintenance. Claude Haiku 5.5 is now the default on the API, offering a 1M context window at $0.10 per million tokens, which lowers the cost floor for high-volume coding tasks. The release also patches critical stability issues in the local agent runtime, specifically fixing memory leaks in HTTP MCP connections and resolving session state corruption during context compaction. These fixes matter because they stabilize the autonomous coding workflow that developers rely on daily. With Haiku 5.5 now standard, teams can deploy cheaper, faster iterations without manual configuration.
Claude Code Releases