vLLM ships 717 commits; DeepSeek V4.1 Flash drops to free
32,788 stars today55 new projects+3,088 ponytail
By olud.ai editorial · built from our own tracking data, published daily
vLLM 0.31.0 arrived on October 5 with 717 commits from 307 contributors, 96 of them new. The release focuses on DeepSeek-V4.1-Flash performance through FlashMLA mega attention and the V4.1 NVFP4 format.
Price moves split sharply. DeepSeek V4.1 Flash input price fell 100% to $0 per 1M tokens, while GLM 5.2 dropped 76%. Several larger models rose, with GLM 5.3 up 536% and DeepSeek V4 Pro 0813 up 150%.
Across the 10,000+ tracked open-source AI projects, 32,788 GitHub stars were gained so far today, 55 new projects entered the index. Project "ponytail" led with +3,088 stars.
This week saw 2,003 new projects added to the index, plus 100 new Hugging Face spaces.
Releases of the day
vLLM 0.31.0 brought 717 commits from 307 contributors, 96 of them new, targeting DeepSeek-V4.1-Flash with FlashMLA mega attention and the V4.1 NVFP4 format. n8n 2.41.7 is a bug-fix release that lets global members grant variable scopes to API keys and shows Gateway credits promotions in the editor. Windmill 1.823.1 fixed Python results with KeyError-raising __getattr__ and now streams the AI chat compaction request. Vespa 8.763.13 updated its command-line tool for local, remote and cloud instances.
Four models saw input price cuts. DeepSeek V4.1 Flash fell 100% to $0 per 1M tokens, GLM 5.2 fell 76% to $0.1, DeepSeek V4 Flash 0423 fell 33% to $0.02, and DeepSeek V3 0324 fell 14% to $0.25. Four rose: GLM 5.3 up 536% to $1.4, DeepSeek V4 Pro 0813 up 150% to $0.55, Llama 3.3 70B Instruct up 120% to $0.22, and Qwen3.8 27B up 2% to $0.43. Two Z.AI models moved in opposite directions.
Two projects entered the index today. "keystone-bench" (Python, 527 stars) is a clinical decision-evidence benchmark that pulls the one fact an answer rests on and checks whether the assistant's action moves with it. "ai-engineering-course" (Markdown, 499 stars) is a free, step-by-step AI engineering course covering machine learning and neural networks.
The Hugging Face community pick, "MotorMind," scaffolds general vision-language models for zero-shot robot manipulation. Vision-language-action models currently generalize poorly to new tasks and environments, and their dependence on specialized training limits their benefit from broader data. MotorMind addresses that without added specialized training.
The day's facts show a widening price split, with one flash model now free while several flagship models rose sharply. Release cadence stays heavy across tools, and ecosystem growth continues at a steady weekly pace.
Source: olud.ai tracking of 10,000+ open-source AI projects, 300+ models and live provider pricing. Figures are measured, not estimated. All releases · Live pricing · Latest in AI
