AI Briefing
KO

DeepSeek V4 Flash now runs on new weights by default on AI Gateway

·2026.07.31 16:00

Key point

DeepSeek V4 Flash now runs by default with new weights on AI Gateway.

Details

DeepSeek V4 Flash now runs by default with updated weights on AI Gateway. Its Terminal-Bench score rose from 56.9 in the April preview to 82.7, a 25.8-point increase, significantly improving agentic task performance.

Users don't need to change their model ID or code — simply calling deepseek/deepseek-v4-flash is enough. Currently only DeepSeek provides the new weights, and additional providers, including others supporting Zero Data Retention, are expected to be supported starting next week.

In the AI SDK, you can specify the model to use it, and coding agents can connect to AI Gateway with the vercel ai-gateway coding-agents setup command and then select DeepSeek V4 Flash in the model settings. AI Gateway applies provider pricing as-is and charges no platform fee on inference and BYOK requests.

deepseekcoding-agentdeepseek-v4-flashai-gatewaybyokai-sdkterminal-bench
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Vercel BlogOpenAI Releases GPT-6 Sol and Luna via AI GatewayVercel BlogAI Gateway Adds TypeSafe Client and HTTP API Support for Jev ModelsHacker NewsDeepSeek Releases DSec: Elastic Sandbox Infrastructure for Large-Scale Agentic TrainingPyTorchKRDeepSeek Releases V4.1-Flash: Replaces V4-Pro and Optimizes KV CacheDeepSeekDeepSeek Releases V4.1-Flash: Outperforms V4-Pro with Lower PricingTossToss Bank Introduces 'Stylepack' Feedback Loop to Resolve Coding Agent Team Rule Violations