DeepSeek prices surge as vLLM adds V4.1 Flash support
86 new projects
By olud.ai editorial · built from our own tracking data, published daily
DeepSeek raised input prices on four model variants today, with increases ranging from 25% to 114%. The largest jump hit DeepSeek V4 Pro 0423, now at $0.90 per 1M tokens, more than double its previous price. Kimi K3 from Moonshot AI rose 76% to $3 per 1M tokens, and Qwen3.8 27B from Alibaba climbed 110% to $0.42 per 1M tokens.
The index added 2,276 new projects this week and 100 new Hugging Face spaces. The ecosystem continues to expand rapidly across both model providers and open-source tooling.
Releases of the day
vLLM v0.30.0 ships with 762 commits from 315 contributors, including 104 first-time contributors. The release adds DeepSeek-V4.1-Flash support with full KV storage, a significant addition given the concurrent price changes from DeepSeek. OpenHands v1.21.0 introduces Docker execution workspace requests from servers, expanding its agent orchestration capabilities. Cline cli-v3.0.64 fixes a critical bug where runs would die when a model turn hit its output-token limit before making a tool call, improving reliability for reasoning-heavy workloads. Pydantic AI v2.47.0 renames the None output route from NoneType to None, a breaking compatibility change. Unstructured 0.27.8 fixes multi-section DOCX partitioning to run in linear time.
Seven models saw price increases today against two decreases. DeepSeek raised four variants: V4 Flash 0423 up 50% to $0.06, V4 Pro 0423 up 114% to $0.90, V4 Pro 0813 up 25% to $0.66. Meta raised Llama 4 Maverick 5% to $0.20. NVIDIA cut Nemotron 3 Nano 30B A3B 17% to $0.05, and Z.AI lowered GLM 5.3 Flash 11% to $0.08. The moves widen the gap between budget and premium tiers.
DeepSeek-V4.1-Flash is now supported in vLLM v0.30.0 with full KV storage, making it available for self-hosted inference deployments immediately.
Research pick
WorldCrafter presents a video world model that uses implicit 3D-aware memory to maintain consistency across long horizons and multiple viewpoints. The model is camera-queryable, meaning users can interactively explore generated environments while the system remembers prior observations.
Today's price increases from DeepSeek and others signal a market correction upward for popular model families, while the vLLM release ensures the open-source community can run these models independently. The research pick points to growing interest in 3D-consistent video generation as a frontier capability.
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
