Who contributed to Local LLMs on August 16, 2026?

15 developers shipped this update, including github-actions[bot], dhiltgen, Fathi Boudra, Eric Zhang, Piotr Wilkin (ilintar), fish-jiang, ngxson, and njhill, and 7 more.

What were the notable Local LLMs updates?

common: migrate the deprecated --mmap/--no-mmap to --load-mode (#26934), common: support --models-dir loading MTP assistant models (#24431), and model: add Kimi-K3 text model (#26185).

Llama.cpp overhauls inference flags, vllm patches deepseek sparse mla · Local LLMs

llama.cpp deprecates mmap flags in favor of unified load-mode, while vllm fixes critical issues blocking DeepSeek-V4-Flash across all decode paths.

llama.cpp's migration away from the deprecated mmap and no-mmap flags marks the end of an era for model loading [1]. The unified load-mode argument now handles all memory management across scripts and examples, simplifying the inference pipeline. Alongside this, the project added support for MTP assistant model discovery through models-dir [2] and introduced Kimi-K3 text model architecture with hybrid KDA/MLA attention, cross-layer residual attention, and latent MoE [3]. On the GPU front, Vulkan got Intel Xe coopmat support via new spec constants and shared memory helpers [4], while the server re-architected its yield_to_queue thread model to fix critical scheduling bugs [5].

vLLM tackled a wave of production blockers across three major systems. The DeepSeek V4 sparse MLA backend now works end-to-end for plain decode, MTP, and DSpark after seven defect fixes verified on Blackwell hardware [6]. The bugfix for DeepSeek V4 eager cudagraph picks the correct region per model runner, restoring performance on ROCm while keeping TTFT gains on MRV2 [7]. Speculative decode keeps EAGLE cache registration on partial-hash-hit paths, fixing a regression that broke fine-grained prefix cache bounds [8]. The frontend added token-aware reasoning token counting to the Streaming Parser Engine with OpenAI-compatible usage fields [9], and EC Connector now builds worker metadata to report per-step state back to the scheduler [10].

SGLang focused on configuration correctness and safety. The config contract now states what resolution actually produced rather than asserting equality, surfacing 297 file/field pairs affected by resolution-written values [11]. Remote media downloads gained exact-hostname allowlisting and streaming size limits to restrict untrusted URLs [12], while a new supply-instance surface pins the contract for raw record inputs [13]. AMD fixes dropped record_stream in MoRI TBO dispatch to prevent HSA out-of-resources hangs [14], and diffusion attention backends now distinguish implicit platform preferences from explicit overrides, letting sparse backends fall back to compatible dense backends [15].

Ollama shipped v0.32.14 with WebP image transcoding for llama-server and Qwen system message tolerance [16], plus an MLX update [17]. Across all four repositories, 64 commits and 64 PRs landed this period with 8 releases.

Local LLMs

LLAMA.CPP OVERHAULS INFERENCE FLAGS, VLLM PATCHES DEEPSEEK SPARSE MLA

LLAMA.CPP SHIPS MODEL OPTIMIZATIONS WHILE VLLM AND SGLANG CHASE MEMORY AND SCHEDULING GAINS

QWEN3.8 LANDS IN OLLAMA WITH DEVELOPER INSTRUCTION SUPPORT

OLLAMA LAUNCHES DEEPSEEK AND MUSE INTEGRATIONS WHILE VLLM SHIPS HARDWARE-AGNOSTIC MODELING

OLLAMA PATCHES CRITICAL SSRF VULNERABILITY IN MANIFEST HANDLING

LANGCHAIN REFRESHES MODEL PROFILES AS OPENHANDS HARDENS LLM VALIDATION

CACHYOS PACKAGES ROLL FORWARD, NANJING MIRROR GOES DARK

ELIXIR CORE SHIPS CALENDAR SPEED SWEEP, PHOENIX LIVE VIEW FIXES UI STALLS

FLUTTER CUTS TEST BOILERPLATE, SWIFT FIXES RUNTIME MEMORY LEAK