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 2026-08-22
stories 266
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

SGLang ships v0.5.18 with 710 PRs and DeepSeek V4 FP4 auto-routing

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

Todd Malsbary Fridge003 ParthSareen hoyyeva jessegross Juan Pérez de Algaba Kevin H. Luu
15 people shipped this
Todd Malsbary @malsbat 2 cited Fridge003 @Fridge003 2 cited ParthSareen @ParthSareen 1 cited hoyyeva @hoyyeva 1 cited jessegross @jessegross 1 cited Juan Pérez de Algaba @jperezdealgaba 1 cited Kevin H. Luu @khluu 1 cited Tiezhen WANG @xianbaoqian 1 cited lukealonso @lukealonso 1 cited Shangming Cai @ShangmingCai 1 cited Aleksander Grygier @allozaur Georgi Gerganov @ggerganov Summer Yang @GirasoleY DarkSharpness @DarkSharpness rainj-me @rainj-me

The biggest release of the week just landed: SGLang v0.5.18 adds new models, fixes a silent disagg bug, and defaults DeepSeek V4 FP4 checkpoints to the right MoE runner.

SGLang shipped v0.5.18 [1], its largest release yet, with 710 PRs from 212 contributors and new model support including Muse Glimmer and Intern-S2-Mobius. In the same period, the team patched a critical bug in disaggregated serving where transfer-failure injection was "silently inert" [2], meaning fault-injection tests were not actually testing anything. For DeepSeek V4 users, FP4 checkpoints now default to the FlashInfer MXFP4 MoE runner on SM90, SM100, and SM120 devices [3], removing a common source of silent performance cliffs. Meanwhile, vLLM made headlines with a back-and-forth on Hunyuan: a PR removed native V1 and VL implementations [4], then a revert restored them [5], so if you use Hunyuan expect churn. vLLM also added a b12x FP4 MoE backend for NVIDIA SM120 and SM121 GPUs [6], a meaningful option for Blackwell users. On the kernel front, llama.cpp landed Q5_K and Q2_K ESIMD kernels [7][8], with measured token-gen improvements up to 1.90x on Q2_K, plus a security validation fix in vLLM's decoder prompt-length checks [9]. Ollama's macOS app now supports Claude model management [10] and a new "Connect your apps" experience [11], while an MLX runner fix addresses the "hanging model" symptom caused by cancelled prefills throwing away computed work [12]. Overall, activity across the four repos totaled 131 commits, 126 PRs, and 9 releases this period.

Action items

References

  1. [1] v0.5.18 ↗ sgl-project/sglang
  2. [2] fix(disagg): PD transfer-failure injection was silently inert (#35890) sgl-project/sglang ↗
  3. [3] [DeepSeek V4] Default FP4 checkpoints to FlashInfer MXFP4 MoE ↗ sgl-project/sglang
  4. [4] Remove native Hunyuan V1 and VL implementations (#53272) vllm-project/vllm ↗
  5. [5] Revert "Remove native Hunyuan V1 and VL implementations" (#53296) ↗ vllm-project/vllm
  6. [6] [Kernel] Add b12x FP4 MoE backend ↗ vllm-project/vllm
  7. [7] Add Q5_K ESIMD kernel ↗ ggml-org/llama.cpp
  8. [8] Add Q2_K reordered MMVQ and ESIMD kernels ↗ ggml-org/llama.cpp
  9. [9] (security) fix: enforce decoder prompt-length validation for skip-che… (#46588) vllm-project/vllm ↗
  10. [10] app: claude model management ↗ ollama/ollama
  11. [11] app: add Connect your apps experience ↗ ollama/ollama
  12. [12] mlxrunner: make prefix cache restore points survive cancelled and resumed prefills ↗ ollama/ollama

$ 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 22, 2026?
The biggest release of the week just landed: SGLang v0.5.18 adds new models, fixes a silent disagg bug, and defaults DeepSeek V4 FP4 checkpoints to the right MoE runner. In total, 131 commits, 126 pull requests, and 9 releases landed.
Who contributed to Local LLMs on August 22, 2026?
15 developers shipped this update, including ParthSareen, hoyyeva, jessegross, Todd Malsbary, Aleksander Grygier, Georgi Gerganov, Juan Pérez de Algaba, and Summer Yang, and 7 more.
What were the notable Local LLMs updates?
v0.5.18, fix(disagg): PD transfer-failure injection was silently inert (#35890), and [DeepSeek V4] Default FP4 checkpoints to FlashInfer MXFP4 MoE.
♥

$ 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.