Update: DeepSeek V4-Flash-Vision-Exp leaps on multimodal agents
An MIT experimental checkpoint competitive with Opus 4.8 on agent tasks, and it serves on vLLM today. Plus: Qwen posts an FP8 build of its Flash MoE.
Copy markdown
First experimental multimodal in the V4 family
DeepSeek posted V4-Flash-Vision-Exp, its “first experimental multimodal model in the DeepSeek-V4 family,” to Hugging Face about six hours ago. The delta over yesterday's V4-Flash-Vision weights: a substantial jump in multimodal agent capability while text-agent performance holds against the V4-Flash-0731 base.
Agent scores that punch at Opus 4.8
The MIT-licensed 305B model reports ApexBench 36.5%, Agents' Last Exam 27.3%, Chartography 64.3% and ZeroBench 35.0% on multimodal tasks, plus Terminal Bench 83.9% and NL2Repo 57.7% on text agents. If those survive your own eval, it's a downloadable agentic vision model in Opus-4.8 territory.
Serve it today on vLLM or SGLang
It ships OpenAI-compatible serving recipes for vLLM and SGLang plus quantized weights (F8_E4M3, I8) for tighter boxes. One caveat DeepSeek states plainly: this is an experimental checkpoint — benchmark it on your own tasks before you wire it into anything that ships.
Elsewhere: Qwen ships an official FP8 Flash-Next
Qwen posted Qwen3.8-Flash-Next-FP8, a block-128 FP8 quant of its 6B-active Flash-Next MoE that it calls “nearly identical” to the BF16 original. It keeps the 262K context (1M via YaRN), and `vllm serve Qwen/Qwen3.8-Flash-Next-FP8` roughly halves the memory footprint.