RepoJournal
Local LLMs Local LLMs
73 wires and counting

$ follow Local LLMs

Keep up with Local LLMs in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-08-10
stories 157
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

OLLAMA SHIPS VISION SUPPORT ACROSS MLX AND QWEN3.5 AS LLAMA.CPP FIXES DISPATCH BUGS

By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology

jessegross github-actions[bot] ch-wan zexplorerhj R3hankhan123 aoshen02 hnyls2002
13 people shipped this
jessegross @jessegross 4 cited github-actions[bot] @github-actions[bot] 2 cited ch-wan @ch-wan 2 cited zexplorerhj @zexplorerhj 1 cited R3hankhan123 @R3hankhan123 1 cited aoshen02 @aoshen02 1 cited hnyls2002 @hnyls2002 1 cited Daniel Hiltgen @dhiltgen allozaur @allozaur yomaytk @yomaytk Caleb DeLeeuw @SolshineCode Srinivasoo7 @Srinivasoo7 thegoldenflow @thegoldenflow

Ollama's MLX runner now processes images natively, preserving prefix caching and speculative decoding while rejecting unsupported requests with clear errors.

Jesse Gross landed the core vision infrastructure for Ollama's MLX backend [1], which expands image tags into placeholder tokens during prepare and routes media through a new media interface that keeps the runner agnostic to model-specific image layouts [4]. Companion work adds Qwen3.5 vision support with interleaved rope positions and causal image expansions that split prefill chunks [3], while a guard prevents models without media capability from crashing on image requests [2]. On the inference engine side, llama.cpp shipped two rapid releases fixing CPU dispatch bugs: the Q5_0 path in SpaceMiT backend [5] and removal of the deprecated HIP ROCWMMA flag [6]. Across 80 commits and 75 PRs in this period, vLLM merged quantization coverage for Ling 3.0 Flash FP8 [7], enabled GPTQ and AWQ on IBM Z s390x architecture [8], and fixed a critical single-node startup race where competing processes could claim rendezvous ports, causing EADDRINUSE failures [9]. SGLang hardened speculative decoding against KV corruption when draft tokens exceed 4 [10] and refactored config reading to respect post-publish overrides across the KV-cache, runner, and scheduler paths [11], [12].

Action items

References

  1. [1] mlxrunner: Add image input support ↗ ollama/ollama
  2. [2] mlxrunner: reject media requests the model cannot serve ↗ ollama/ollama
  3. [3] qwen3_5: image input support ↗ ollama/ollama
  4. [4] mlxrunner: expand image tags into placeholder tokens ↗ ollama/ollama
  5. [5] b10333 ↗ ggml-org/llama.cpp
  6. [6] b10332 ↗ ggml-org/llama.cpp
  7. [7] `[Model][Quantization] Add Ling-3.0-flash-fp8 support` ↗ vllm-project/vllm
  8. [8] [CPU] Enable GPTQ and AWQ quantization for s390x ↗ vllm-project/vllm
  9. [9] [BugFix] Use file:// rendezvous for single-node executors to eliminate startup port races ↗ vllm-project/vllm
  10. [10] [DSV4] Fix silent KV corruption when speculative draft tokens > 4 ↗ sgl-project/sglang
  11. [11] config: the KV-cache configurator reads the bags ↗ sgl-project/sglang
  12. [12] config: the runner and scheduler read resolved config from the bags ↗ sgl-project/sglang

$ ls local-llm/reviews/ # the step-back reads

the week in review vLLM fixes cache_salt oracle and multimodal IPC desync the month in review vLLM 0.30.0 lands and nineteen engine-fatal flaws close
Full archive → Weekly reviews →

Quick answers

What shipped in Local LLMs on August 10, 2026?
Ollama's MLX runner now processes images natively, preserving prefix caching and speculative decoding while rejecting unsupported requests with clear errors. In total, 80 commits, 75 pull requests, and 2 releases landed.
Who contributed to Local LLMs on August 10, 2026?
13 developers shipped this update, including jessegross, Daniel Hiltgen, github-actions[bot], allozaur, yomaytk, Caleb DeLeeuw, zexplorerhj, and R3hankhan123, and 5 more.
What were the notable Local LLMs updates?
mlxrunner: Add image input support, mlxrunner: reject media requests the model cannot serve, and qwen3_5: image input support.
♥

$ log Local LLMs --recent

OpenVINO 2026.4.1 lands in llama.cpp, vLLM frees CUDA graph memory on sleep 156 shipped llama.cpp spec sampling goes probabilistic, vLLM trims GLM-5.3 attention overhead 262 shipped Qwen4Exp MTP lands in llama.cpp as SGLang ships v0.5.21 281 shipped MLX ports Metal gather_qmm wins to CUDA, heads-up on a qmm_sm80 corruption bug 330 shipped

$ log --all # elsewhere on the wire

@vuejs Vue language tools stop calling registerHooks on buggy Node builds @openclaw OpenClaw's deslop sweep removes 1,000+ production lines

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.