RepoJournal
Local LLMs Local LLMs
93 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 2026-08-11
stories 271
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

OLLAMA PATCHES CRITICAL PARSER BUG, NEMOTRON 3 LANDS WITH SPECULATIVE DECODING

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

dhiltgen ch-wan Pedro Cuenca superm1 gaugarg-nv khluu Priyjain-amd
15 people shipped this
dhiltgen @dhiltgen 3 cited ch-wan @ch-wan 2 cited Pedro Cuenca @pcuenca 1 cited superm1 @superm1 1 cited gaugarg-nv @gaugarg-nv 1 cited khluu @khluu 1 cited Priyjain-amd @Priyjain-amd 1 cited AndreasKaratzas @AndreasKaratzas 1 cited mmangkad @mmangkad 1 cited Pascal @ServeurpersoCom github-actions[bot] @github-actions[bot] aoshen02 @aoshen02 BugenZhao @BugenZhao Brayden Zhong @b8zhong Khoa Pham @kpham-sgl

Ollama shipped v0.32.7 with a critical fix for boundary token corruption in model invocation, while full MLX support for Nemotron 3 Nano Omni brings self-draft speculation to inference without a separate draft model.

The parser bug [1] hit production when models emit <|message|> boundary tokens in positions that broke ATEM parameter parsing, replacing the expected terminator entirely. The fix replaces strip-after-cut recovery with a unified name-scan path that handles both the terminated and unterminated shapes. In parallel, Nemotron 3 support [2] lands with Metal-optimized quantized expert kernels and a built-in multi-token prediction head that serves as a self-draft speculator, eliminating the need for a separate draft model during speculative decoding. llama.cpp moved fast this cycle: Muse Glimmer architecture support [4] shipped, ROCm 7.14 CI targets [5] are now live, and backend sampling extended from single to multiple outputs per sequence [6], unblocking speculative decoding in the inference engine. vLLM v0.27.0 [7] landed with Kimi K3 full-stack support across core, kernels, Python and Rust frontends, plus CPU Zen MoE routing through zentorch [8] ahead of fallback GEMM paths. SGLang hardened spec-decode testing [9], fixed DSpark shared-expert layout with the DSV4 gate [10], and unified memory now works with PD disaggregation [11] after fixing a NaN logits bug in deterministic Triton extend [12].

Action items

References

  1. [1] parsers: recover boundary tokens fumbled into glimmer ATEM invoke names ↗ ollama/ollama
  2. [2] mlx: implement Nemotron 3 ↗ ollama/ollama
  3. [3] Release v0.32.7 ↗ ollama/ollama
  4. [4] model: Muse Glimmer Support (#26841) ↗ ggml-org/llama.cpp
  5. [5] Add CI targets for ROCm 7.14 ↗ ggml-org/llama.cpp
  6. [6] Backend sampling multi output ↗ ggml-org/llama.cpp
  7. [7] v0.27.0 ↗ vllm-project/vllm
  8. [8] [CPU][Zen] Route BF16 MoE inference through zentorch on AMD ↗ vllm-project/vllm
  9. [9] [CI] Solidify speculative decoding E2E coverage ↗ vllm-project/vllm
  10. [10] Fix DSV4 DSpark shared expert loading ↗ sgl-project/sglang
  11. [11] [PD] Support --enable-unified-memory with PD disaggregation (kimi-linear MLA hybrid-Mamba) ↗ sgl-project/sglang
  12. [12] Fix NaN logits from deterministic Triton extend on the unified memory pool ↗ sgl-project/sglang

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

the week in review vLLM v0.32.0 lands systemone, security fixes behind it 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 11, 2026?
Ollama shipped v0.32.7 with a critical fix for boundary token corruption in model invocation, while full MLX support for Nemotron 3 Nano Omni brings self-draft speculation to inference without a separate draft model. In total, 129 commits, 129 pull requests, and 13 releases landed.
Who contributed to Local LLMs on August 11, 2026?
15 developers shipped this update, including dhiltgen, Pedro Cuenca, Pascal, github-actions[bot], superm1, gaugarg-nv, khluu, and Priyjain-amd, and 7 more.
What were the notable Local LLMs updates?
parsers: recover boundary tokens fumbled into glimmer ATEM invoke names, mlx: implement Nemotron 3, and Release v0.32.7.
♥

$ log Local LLMs --recent

Ollama preps Phonon voice input, vLLM drops deprecated IR knob 269 shipped Ollama fixes cloud metadata for every launcher, MLX speeds 6-16 row decode ~3x 298 shipped Ollama streams real errors, llama.cpp forwards sampling params on /v1/messages 274 shipped Five vLLM crash and forgery paths, realtime transcription lands in llama.cpp 295 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training MILES ships a 25-benchmark eval panel and simpler config presets Interpretability J-space direction loss and Jacobian lens joint ablation fixed in v4.3.0

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Interpretability Alignment & Post-training

$ 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.