AI Pulse · AI Trends Pulse

DeepSeek just pushed V4-Flash-0731 into public beta and released the model weights at the same time. This is a sparse mixture-of-experts architecture with 256 routed experts, activating six at a time per token. It handles up to one million tokens of context and ships with a Codex-compatible API, so existing code tooling can plug straight in. Same-day serving support landed for vLLM and SGLang, which means you can spin it up on your own infrastructure without waiting on vendor integrations.

The sparse routing matters because it keeps inference costs down. You’re not running all 256 experts for every token, just the six the model decides are relevant. That makes a model this size cheaper to serve than a dense equivalent, especially at scale. The million-token context window is useful if you’re feeding it entire codebases, long documents, or multi-turn conversations that need to remember everything. The Codex API compatibility means if you’ve already built workflows around OpenAI’s code models, you can swap this in with minimal rewrite.

Public beta plus open weights is the real signal. You can test it in production without negotiating enterprise contracts, and you can run it on your own hardware if you don’t want to send data outside your network. The same-day vLLM and SGLang support tells you the serving ecosystem is ready to handle it, which removes a common deployment bottleneck. If you’re running AI workloads that involve code generation, large document reasoning, or anything that chews through tokens fast, this is worth benchmarking against whatever you’re using now. We build these kinds of deployment options into the Omni Command Centre so teams can test and switch models without rearchitecting their stack every time a new one drops.