News · 2026-09-06

DeepSeek releases a 168 GB MIT-licensed multimodal V4 checkpoint

DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an MIT-licensed experimental multimodal checkpoint whose Hugging Face repository totals 168 GB. The release matters because it makes a large vision-capable agent model downloadable under a permissive license, while also showing why release claims need careful reading: its headline comparisons are DeepSeek's own measurements under DeepSeek's own harness.

Key facts

• DeepSeek's API changelog dates the release to August 21, 2026.
• The model card calls it DeepSeek's first experimental multimodal model in the V4 family.
• The repository lists an MIT license and 168 GB of files.
• The primary materials state no minimum or recommended VRAM requirement; disk size is not a safe proxy for runtime memory.

The model adds visual understanding to DeepSeek-V4-Flash. DeepSeek positions it for multimodal agent work: tasks where a system inspects an interface, chart, image, or visual state and then acts through tools. The company reports 83.9 on Terminal Bench 2.1, 57.7 on NL2Repo, 59.3 on DeepSWE, 64.3 on Chartography, and other results. It says the model is 'on par' with V4-Flash for text-agent tasks and 'close to Opus-4.8' on multimodal-agent benchmarks.

Those statements are verified as what DeepSeek reports, not as independent conclusions. The same card specifies DeepSeek Harness minimal mode, maximum reasoning effort, temperature 1.0, and top_p 0.95 for its text-agent benchmarks. Harness choices determine prompts, tools, retries, and scoring. It is like comparing race cars after one manufacturer also chooses the tires and track conditions. A benchmark number can be meaningful without proving a vendor-neutral ranking.

The operational fact is the release format. A reader can obtain a 168 GB checkpoint rather than depend only on an API, enabling private deployment, inspection, and research. It also creates a material operating constraint. A 168 GB download is not a 168 GB VRAM requirement: weights must be loaded, inference needs activations and KV cache, and implementations may use different precisions and sharding. The card does not make a primary-source memory recommendation, so claiming it runs on a particular GPU would be speculation.

This is why open weights require more than a label. License, files, documentation, hardware, and operational dependencies determine practical openness. DeepSeek has made the legal and download layer unusually clear. Teams still need to inspect the code path, configuration, safety controls, and infrastructure before serving the model. The newsworthy fact is not that DeepSeek has settled the leaderboard; it is that it shipped a large, permissively licensed vision-agent checkpoint with concrete benchmark disclosures and an observable 168 GB distribution footprint.

Primary source, verified: read the paper →

Ground Truth. (2026, September 6). DeepSeek releases a 168 GB MIT-licensed multimodal V4 checkpoint. Ground Truth. https://groundtruth.day/news/deepseek-v4-flash-vision-exp-releases-mit-licensed-168gb-checkpoint.html

@misc{groundtruth:deepseek-v4-flash-vision-exp-releases-mit-licensed-168gb-checkpoint,
  title  = {DeepSeek releases a 168 GB MIT-licensed multimodal V4 checkpoint},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/deepseek-v4-flash-vision-exp-releases-mit-licensed-168gb-checkpoint.html}
}

Topics: deepseek · open-weights · multimodal · agents · models