Blog · Signals · September 21, 2026
Signals: what shipped in local AI, week of September 21, 2026
14 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.
-
vLLMv0.30.0: [Build] Fix DeepGEMM CUDA 12.9 release builds (#57554)2026-09-21
Signed-off-by: khluu khluu000@gmail.com
-
Ollamav0.34.32026-09-19
What's Changed GET /api/show now advertises each model's thinking controls and default: curl -d ' {"model": "glm-5.3-flash:cloud"} ' { "thinking" : { "values" : [ " low " , " high " , " max " ],…
-
vLLMv0.30.0rc22026-09-18
[Bugfix][NIXL] Avoid receive reports for notification-only requests (…
-
Ollamav0.34.3-rc02026-09-18
api: expose model thinking levels and defaults
-
FreeTokenv0.1.32026-09-16
Highlights Image input. Qwen3-VL, Qwen3.6, Qwen3.8-Flash-Next, Gemma-4, GLM-5.3-Flash, MiniMax-M3 and Muse-Glimmer-30B accept images through the OpenAI, Anthropic and Responses APIs. Vision towers…
Models
Open weights published in the open, newest first.
-
Trending GGUFabenzerps/Qwen-Image-2.1-GGUF2026-09-20
33,232 downloads · 368 likes · gguf
-
QwenQwen/Qwen-Image-2.1-PE-I2I2026-09-20
4,604 downloads · 31 likes
-
QwenQwen/Qwen-Image-2.1-PE-T2I2026-09-20
876 downloads · 30 likes
-
Trending GGUFdealignai/Bonsai-2-27B-Ternary-CRACK-GGUF2026-09-17
33,051 downloads · 113 likes · gguf · text-generation
-
NVIDIA modelsnvidia/DeepSeek-V4.1-Flash-NVFP42026-09-16
2,413 downloads · 80 likes
-
NVIDIA modelsnvidia/c-foundationstereo-s2026-09-15
180 downloads · 7 likes
Hardware and platform
The silicon and the runtime everything above depends on.
-
Hugging Face blogPruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem2026-09-21
-
Node.js releases2026-09-16, Version 26.9.0 (Current), @aduh952026-09-16
Notable Changes [ d414624dce ] - (SEMVER-MINOR) crypto : add a generic MAC API (Filip Skokan) #65553 [ 7ac458f802 ] - (SEMVER-MINOR) crypto : discover ciphers from OpenSSL providers (Filip Skokan)…
-
NVIDIA blogNVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut2026-09-16
System performance, efficient infrastructure scaling and continuous software optimization are key levers that determine AI inference economics. Higher system performance means more tokens generated,…
Sources
Every entry above links its own announcement. The feeds read this week:
- vLLM
- Ollama
- FreeToken
- Trending GGUF
- Qwen
- NVIDIA models
- Hugging Face blog
- Node.js releases
- 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.