7 noviniek
DeepSeek open-sourced DeepInfer, the production inference framework powering its large-scale services, supporting DeepSeek-V4.1-Flash with speculative decoding (DSpark), deterministic sampling, and structured outputs on NVIDIA SM100 and Huawei Ascend 950.
DeepSeek V4 graduated from preview to general availability with a 1M-token context window, V4-Pro and V4-Flash variants, MIT-licensed open weights, and the AI industry's first structural peak/off-peak API surge pricing.
Bolt.new introduced Skills -- reusable bundles of context, rules, and workflows (tech stack, design standards, code review rules) that Bolt applies during development, shareable across individual projects or entire workspaces.
DeepSeek introduced peak-hour pricing for its V4 API — token rates double during Beijing business hours (9–12 and 14–18 CST). The deepseek-chat alias retires July 24; migration to deepseek-v4-flash or deepseek-v4-pro is mandatory.
DeepSeek will deactivate legacy aliases deepseek-chat and deepseek-reasoner on July 24, 2026 at 15:59 UTC. Developers must update calls to deepseek-v4-pro or deepseek-v4-flash — note critical mapping caveat for deepseek-reasoner.
DeepSeek's official V4 release is coming mid-July with a 1-million-token context window and the platform's first-ever peak/off-peak API pricing — calls during Beijing business hours will cost 2x the off-peak rate.
Qwen-AgentWorld is the first model trained to predict the next environment state (not the next action) across seven agent domains. Two MoE sizes, 256K context, Apache 2.0.
