RepoJournal
Local LLMs Local LLMs
46 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 2026-09-03
stories 258
RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

Error propagation hardened in ollama's MLX bindings, gemma4 gains multimodal support

By RepoJournal · Filed · About Local LLMs

jessegross dhiltgen ngxson am17an Georgi Gerganov Adrien Gallouët Kevin H. Luu
15 people shipped this
jessegross @jessegross 4 cited dhiltgen @dhiltgen 1 cited ngxson @ngxson am17an @am17an Georgi Gerganov @ggerganov Adrien Gallouët @angt Kevin H. Luu @khluu yzeyu71 @yzeyu71 AndreasKaratzas @AndreasKaratzas NickLucche @NickLucche akhauriyash @akhauriyash merrymercy @merrymercy oulgen @oulgen JustinTong0323 @JustinTong0323 ByronHsu @ByronHsu

ollama's MLX runner now treats every fallible MLX call as an error it can surface, ending silent failures that corrupted outputs.

Ollama's MLX bindings dropped a process-exiting default error handler and then ignored the return codes, so failed ops continued with a null array and surfaced later as zeros, skipped evals, or crashes in unrelated code. The new [1] and [2] work checks every fallible call, panics with the captured MLX message, and returns load/save errors directly, so failures fail loudly at the point of origin. This is a correctness fix for anyone running gemma4 or other safetensors models on Apple Silicon via the MLX engine.

Gemma4 imports served by the MLX engine now accept image and audio input. Images run through the transformer tower (26B, 31B, e-series) and the 12B's encoder-free unified embedder, while audio flows through WAV bytes in the images field, OpenAI input_audio, and /v1/audio/transcriptions. The server no longer hides these capabilities [3]. This brings MLX parity with the GGUF-side multimodal support [4] already present in ollama.

Cache-aware metrics ship: native responses gain prompt_eval_cached_count, and OpenAI- and Anthropic-compatible APIs expose equivalent fields, keeping prompt_eval_count as the logical total while excluding cache hits from prefill rate calculations. llama-server and MLX both feed the counts [5].

Action items

References

  1. [1] mlxrunner: capture MLX errors in a single buffer read after every call ↗ ollama/ollama
  2. [2] mlxrunner: check every fallible mlx-c call in the MLX bindings ↗ ollama/ollama
  3. [3] gemma4: image and audio input on the MLX engine ↗ ollama/ollama
  4. [4] gemma4: image and audio input support ↗ ollama/ollama
  5. [5] Report cached prompt tokens ↗ ollama/ollama

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

the week in review Vulkan top_k radix select arrives for long rows the month in review Speculative decoding consolidates across the stack as DeepSeek V4 lands
Topics: Local LLMs AI / ML Full archive →

Quick answers

What shipped in Local LLMs on September 3, 2026?
ollama's MLX runner now treats every fallible MLX call as an error it can surface, ending silent failures that corrupted outputs. In total, 125 commits, 123 pull requests, and 10 releases landed.
Who contributed to Local LLMs on September 3, 2026?
15 developers shipped this update, including jessegross, dhiltgen, ngxson, am17an, Georgi Gerganov, Adrien Gallouët, Kevin H. Luu, and yzeyu71, and 7 more.
What were the notable Local LLMs updates?
mlxrunner: capture MLX errors in a single buffer read after every call, mlxrunner: check every fallible mlx-c call in the MLX bindings, and gemma4: image and audio input on the MLX engine.
♥

$ log Local LLMs --recent

Vulkan adds TQ1_0, Spark2.5 lands in llama.cpp 180 shipped vLLM closes scale-out multimodal handoff hole 86 shipped vLLM Fast Start maps weights zero-copy; GLMGA sampling capped 256 shipped Ollama enables speculative decoding under structured output, vLLM adds LoRA and fusion 259 shipped

$ log --all # elsewhere on the wire

Meta React Native adds Stable API guards to renderer attributed string FastAPI & Pydantic Pydantic fixes Mypy plugin crash, tightens datetime constraints

$ 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.