RepoJournal
Local LLMs Local LLMs
99 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 2026-10-28
stories 282
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

LFM3 lands across llama.cpp and Ollama, vLLM fields four security fixes

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

0cc4m tdakhran Juan Pérez de Algaba jlqibm dhiltgen tc-mb ggerganov
14 people shipped this
0cc4m @0cc4m 1 cited tdakhran @tdakhran 1 cited Juan Pérez de Algaba @jperezdealgaba 1 cited jlqibm @jlqibm 1 cited dhiltgen @dhiltgen tc-mb @tc-mb ggerganov @ggerganov Aleksander Grygier @allozaur BBuf @BBuf metamergebot @metamergebot AgainstEntropy @AgainstEntropy chuyeh @chuyeh zcbenz @zcbenz StockerMC @StockerMC

LiquidAI's LFM3 family picks up first-class support in both llama.cpp and Ollama on the same day vLLM ships four security patches, three of them unauthenticated attack surfaces.

model : add LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B ggml-org/llama.cpp

by tdakhran

llama.cpp adds LiquidAI/LFM3-1.7B and LFM3-VL-2B, with video input deferred and correctness checked against the reference implementation. The companion Ollama change honors an explicit tokenizer.ggml.add_bos_token, since LFM3 GGUFs set it to false and the lfm2 default was stripping the template's BOS; on lfm2:24b, WikiText PPL drops from 81.3 to 72.0.

[Security] Fix ReDoS in validation error path sanitizer (#52120) vllm-project/vllm

by Juan Pérez de Algaba

vLLM patches a ReDoS in the validation error path sanitizer, one of four security fixes landing here. The other three: a sanitize_message() address regex that only matched lowercase and so failed to strip PIL's uppercase __repr__ heap addresses from responses to unauthenticated clients, a DoS via prompt embeds on XD-RoPE models that is an incomplete fix of CVE-2026-55514, and quadratic Cohere2 ...

sanitize_message() address regex is lowercase-only and does not match PIL's uppercase __repr__, leaking live heap addresses to unauthenticated clients vllm-project/vllm

The address leak is the one to size carefully: sanitize_message() was the scrubber between PIL exception text and the client, and it missed the uppercase form, so live heap addresses reached callers who never authenticated.

server: add router device memory margin parameter for dynamic unloading ggml-org/llama.cpp

by 0cc4m

llama.cpp's server adds --models-memory-margin, letting the router keep several small embedding models resident and unload them when a large text model needs the space. It behaves like the autofit memory margin, tracking per-model requirements and evicting in the existing max-number order.

[Fix] compressed-tensors: Serve split-expert MXFP4 MoE via FP4 Marlin sgl-project/sglang

by jlqibm

A compressed-tensors mxfp4-pack-quantized MoE checkpoint with experts split per expert (experts.{e}.{gate,up,down}_proj.*, as llmcompressor emits) now loads through FP4 Marlin in SGLang; before this it failed with a KeyError on w2_weight_packed. The earlier fix covered MXFP4 dense linear only. Elsewhere: MOSS-Image packs its joint-block QKV projections for quality=high, FlashAttention calls rou...

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

the week in review Three vLLM breaking issues expose default-server attack paths 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 October 28, 2026?
LiquidAI's LFM3 family picks up first-class support in both llama.cpp and Ollama on the same day vLLM ships four security patches, three of them unauthenticated attack surfaces. In total, 133 commits, 133 pull requests, 10 releases, and 6 security advisories landed.
Who contributed to Local LLMs on October 28, 2026?
14 developers shipped this update, including dhiltgen, tc-mb, ggerganov, 0cc4m, Aleksander Grygier, tdakhran, Juan Pérez de Algaba, and BBuf, and 6 more.
What were the notable Local LLMs updates?
model : add LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B, [Security] Fix ReDoS in validation error path sanitizer (#52120), and sanitize_message() address regex is lowercase-only and does not match PIL's uppercase __repr__, leaking live heap addresses to unauthenticated clients.
♥

$ log Local LLMs --recent

Faster Metal MXFP4 decode, a GLM-4.7 streaming parser stall fixed 275 shipped llama.cpp lands SYCL DPAS flash attention, Ollama repairs Responses history 289 shipped Ollama fixes OpenCode v2 launch, llama.cpp rewrites CPU concat as one kernel 297 shipped Metal grid truncation fixed, vLLM patches GLM-5.3 weight loss 236 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training XPU rollout images, a tensordict 0.11 data fix, and a gated-judge guard Agent Security Denied RAG queries get real hashes, Docker deployer logs gain stderr

$ 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.