Blog · Signals · October 5, 2026
Signals: what shipped in local AI, week of October 5, 2026
12 things shipped in local AI this week that we thought were worth knowing about. Every line below is a project's own announcement of its own work, linked back to where it was published.
What this page is. A machine collected these from release feeds and model registries and formatted them. It did not summarize, rank by opinion, or add commentary — the words in each entry are the publisher's own words, and the date is theirs too. That is the whole reason it is allowed to publish itself: there is nothing here for a machine to get wrong except the copying, and the link beside every line is how you check it. Inclusion is not endorsement, and nothing here has been tested by us. The pages where we do make claims about our own product are written by people and reviewed, and our own limits stay recorded in the Sovereignty Ledger.
Engines
The local runtimes SovereignAI talks to, and the ones it could.
-
llama.cppb11422: cuda: use the vector lightning indexer kernel on MUSA (#29990)2026-10-05
cuda: stage the lightning indexer queries in head passes for MUSA MUSA archs 21 and 22 cap static shared memory at 28 KB, and the tile kernel staged the queries of all four heads next to the key tile…
-
vLLMv0.31.02026-10-05
v0.31.0 Highlights This release features 717 commits from 307 contributors (96 new)! DeepSeek-V4.1-Flash performance : FlashMLA mega attention with the V4.1 NVFP4 compressed KV cache is now the SM100…
-
Ollamav0.40.02026-10-05
What's Changed Models run on MLX on Apple Silicon by default In this release, on Apple Silicon devices, model architectures supported by the MLX runtime will automatically run on MLX. ollama pull…
-
Ollamav0.40.0-rc1: mlx: match publisher tokenizer semantics (#18779)2026-10-04
mlx: match publisher tokenizer semantics Honor pretokenizer stage order, split behavior, Unicode boundaries, added-token normalization, and ranked BPE merges. Handle empty added tokens and empty…
-
vLLMv0.31.0rc42026-10-02
[Bugfix][HiSparse] Fix MTP acceptance collapse under FULL graphs with…
Models
Open weights published in the open, newest first.
-
Google open weightsgoogle/DiarizationLM-Gemma-4-E4B-v12026-10-04
880 downloads · 24 likes · gguf
-
NVIDIA modelsnvidia/PixelDiT2-ImageNet2026-10-01
10 likes
-
NVIDIA modelsnvidia/PixelUMM2026-10-01
57 likes
-
Trending GGUFISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF2026-09-26
389,398 downloads · 270 likes · gguf · moe
-
Trending GGUFViggle/Qwen-Image-2.1-viggle-turbo2026-09-22
286,885 downloads · 601 likes · gguf
Hardware and platform
The silicon and the runtime everything above depends on.
-
NVIDIA blogNVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI2026-10-02
Local AI is becoming more useful by the token. As AI agents move from experiments into everyday development, increasingly capable open models are shrinking to fit on more devices, giving builders…
-
NVIDIA blogHow NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast2026-10-01
GPT-6 Astra Ultrafast, running on NVIDIA Blackwell GPUs, is available now in the OpenAI API and to eligible ChatGPT Work and Codex users. Accelerated by inference optimizations through OpenAI’s…
Sources
Every entry above links its own announcement. The feeds read this week:
- llama.cpp
- vLLM
- Ollama
- Google open weights
- NVIDIA models
- Trending GGUF
- NVIDIA blog
The full source list, with a note on why each one is watched, is in the repository — as is the code that assembled this page. Nothing was fetched that is not in that file.