Who contributed to Local LLMs on July 25, 2026?

15 developers shipped this update, including dhiltgen, ParthSareen, Xuan-Son Nguyen, Johannes Gäßler, Pascal, Kevin H. Luu, Andreas Karatzas, and Harry Mellor, and 7 more.

What were the notable Local LLMs updates?

model: add Laguna MLX support, mlx: keep loaded model memory resident, and server: fix ps data race on scheduler loaded map (#17376).

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.

Ollama's primary push lands Laguna XS 2 and 2.1 support with mixed-precision quantization across MLX's dense and routed MoE layers [1], paired with a Metal residency optimization that keeps loaded models memory-resident to unlock the performance gains those new quantization policies promise [2]. The same release fixes a data race in the process handler that was iterating the scheduler's loaded map without locks, allowing concurrent unload operations to nil model references mid-request, which would corrupt state under load [3]. On the agent side, Ollama added system prompt inspection to the TUI and permission gating for skill tool invocation, requiring explicit approval when the model attempts to load skills but preserving user-initiated skill activation without friction [4] [5]. SGL-Lang v0.5.16 ships DSpark, a new speculative algorithm that "drafts semi-autoregressively in blocks, then sizes each verify window from the draft's own confidence instead of a fixed draft length" [6], reaching 383.7 tok/s at accept length ~5 on DeepSeek-V4-Pro with tensor parallel 8 on B300. The release also wires DSV4's indexer topk transform through the DSA backend for deterministic RL rollouts, adds prefill and decode load counters to LoadSnapshot for deeper scheduling telemetry, and upgrades audio decode to handle browser MediaRecorder output and mislabeled MIME types [7] [8] [9]. LLaMA.cpp removed rocWMMA FlashAttention from HIP and render effects from the web UI that were leaving content invisible when blocks mounted offscreen during long agentic loops, while fixing settings precedence so your UI choices always override admin config [10] [11]. vLLM's push-mode distributed pipeline fixed a blocking handshake call on the writer thread that was causing noticeable performance degradation when new decoder instances scaled up, deferring registration calls to async paths like the pull-mode already did [12].