53,670 stars today107 new projects+4,161 rocketride-server
By olud.ai editorial · built from our own tracking data, published daily
DeepSeek raised input prices across four models on September 29, 2026, with V4 Pro 0423 up 123 percent to $0.78 per 1M tokens. V3.2 rose 4 percent to $0.28, V3 0324 rose 16 percent to $0.29, and V3.2 Exp fell 52 percent to $0.13. The moves come as the ecosystem gained 53,670 GitHub stars across 10,000+ tracked projects, with 107 new projects entering the index.
Weekly context shows 3,078 new projects joined the index this week, plus 100 new Hugging Face spaces. Rocketride-server leads GitHub today with +4,161 stars. The pricing shifts hit a broad set of providers, with MiniMax M2.7 up 43 percent to $0.30 and Google's Gemma 4 26B A4B up 14 percent to $0.08.
Releases of the day
RAGFlow v1.0.0-rc1 marks a comprehensive rewrite in Go, promising a substantial reduction in system resource usage. Langfuse v4.47.0 adds dedicated trace transcripts for topics. Docling v2.131.0 introduces Apple Keynote support and parses Keynote charts as classified pictures. Windmill v1.819.0 adds a hub sync button for resource types and provenance claims to job OIDC tokens. MinerU 4.0.10 fixes WebUI UUID fallback in non-secure contexts. Cline sdk v0.0.87 fixes hub startup on Windows machines with system proxies.
Moonshot AI cut Kimi K2.6 input price 32 percent to $0.65 per 1M tokens. Alibaba dropped Qwen3.5-35B-A3B 48 percent to $0.16. DeepSeek V3.2 Exp fell 52 percent to $0.13. The increases cluster around DeepSeek and MiniMax, while the decreases target mid-range and experimental models.
▲ 123%DeepSeek V4 Pro 0423DeepSeek · input $0.35 → $0.78 per 1M tokens
Swarm-forge, a Clojure tool for coordinating several AI agents, leads new detections with 3,930 stars. ClawBench, an open-source benchmark for browser AI agents, gained 882 stars. AI-Visual-Prompt-Cookbook offers 150+ JSON style prompts for image models with 681 stars. Bello, a Python guard for long-horizon coding tasks, has 86 stars.
A Hugging Face community paper with 58 upvotes proposes domain-normalized multi-teacher on-policy distillation. The method trains one model with skills from several specialist teachers, each excelling at a single task like math or coding, avoiding the need for users to switch between models.
Today's price moves show providers repricing aggressively, with DeepSeek raising flagship costs while cutting experimental variants. The RAGFlow Go rewrite signals a push toward lighter infrastructure. New agent-coordination tools and benchmarks suggest continued focus on multi-agent reliability.
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
