DeepSeek cuts flash model prices; DSPy 3.4 and Pydantic AI release
23,788 stars today101 new projects+1,161 paperclip
By olud.ai editorial · built from our own tracking data, published daily
DeepSeek cut input prices on three of its V4 models, with the flash variant dropping 25% to $0.03 per million tokens. The price reductions come alongside a busy release day for open-source AI tools.
The ecosystem gained 23,788 GitHub stars today across the 10,000 tracked projects, and 101 new projects were added to the index. Paperclip led daily gains with +1,161 stars.
This week, 2,371 new projects entered the index and 100 new Hugging Face spaces were created. The pace of new tooling and model releases continues to accelerate.
Releases of the day
n8n shipped version 2.40.7 with a fix that propagates project span attributes to node spans, improving observability in workflows. DSPy 3.4 introduced Jev integration via TypeSafe, adding three experimental decision types that bring probability evidence into ordinary prompts: Noul, Choice, and Score. Pydantic AI v2.50.0 now lets developers ask a DecisionModel which route to take by name, with one label per route. TRL v1.14.0 removed its trl.losses module — DPO, KTO and GRPO now stream their own log-probs directly. LoopX v1.2.0 targets operators using Goal Chat, steward routing, and team delegation, adding verifiable team work and recoverable local agents.
Input price cuts hit multiple models. DeepSeek V4 Flash 0423 fell 14% to $0.06 per million tokens, V4 Pro 0423 fell 9% to $0.86, and V4 Flash 0731 fell 25% to $0.03. Kimi K2.7 Code from Moonshot AI dropped 7% to $0.66. Several models rose: Qwen3 Next 80B A3B Instruct climbed 11% to $0.1, Qwen3 30B A3B Instruct 2507 doubled to $0.1, MiniMax M2 rose 15% to $0.3, and GLM 5.3 from Z.AI jumped 67% to $1.4.
Two projects entered the index today. Chubbyskills (966 stars) collects Chinese content from multiple channels into a personal knowledge base using 13 AI skills. Deja-vu (974 stars) provides a single memory shared by 31 coding agents including Claude Code, Codex, and Cursor, built from session history.
A Hugging Face community paper with 57 upvotes, "Training Object Permanence in World Models" by Haotian Zhang, Fengyuan Yu, Dezhi Luo and others, explores how video generation models can learn the human cognitive priors of object permanence and solidity. The authors argue that world models showing reasoning abilities can be further improved by explicitly training for object persistence.
Today's facts point to an increasingly competitive model pricing landscape, led by DeepSeek's aggressive cuts on flash-tier models. The steady stream of agent-framework releases and the emergence of memory-sharing tools indicate that the ecosystem is maturing beyond single-agent paradigms. Research into cognitive priors for world models suggests the next frontier may be in imbuing generative models with core human intuitions.
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
