Superpower Daily
Caterpillar Commits $100M to Retrain 118,000 Employees as AI Mo...Broadcom Releases AgentMinder to Check AI Agent Tool Calls at R...DeepSeek Releases 305B Vision Weights Under MIT License, but Lo...Claims Adjusters’ AI Backlash Hits 98% Negative Reviews as Insu...Deloitte Buys Wavicle Assets to Add Databricks Engineering to I...Replica Cyber Ships Isolated AI Agents With Runtime Controls fo...Caterpillar Commits $100M to Retrain 118,000 Employees as AI Mo...Broadcom Releases AgentMinder to Check AI Agent Tool Calls at R...DeepSeek Releases 305B Vision Weights Under MIT License, but Lo...Claims Adjusters’ AI Backlash Hits 98% Negative Reviews as Insu...Deloitte Buys Wavicle Assets to Add Databricks Engineering to I...Replica Cyber Ships Isolated AI Agents With Runtime Controls fo...
Modelspublished

DeepSeek Releases 305B Vision Weights Under MIT License, but Local Use Needs Serious Hardware

The release gives developers the model files and serving references needed to run or adapt DeepSeek-V4-Flash-Vision-Exp independently. Its 168GB size means that freedom is aimed more at infrastructure-equipped teams than typical individual users.

By 3 min read
DeepSeek Releases 305B Vision Weights Under MIT License, but Local Use Needs Serious Hardware
DeepSeek Releases 305B Vision Weights Under MIT License, but Local Use Needs Serious Hardware
Sign in to save and react

Listen to this story

The audio brief

About 1:34
0:001:34
Read transcript
DeepSeek has released the downloadable weights for its experimental V4 Flash Vision model, putting a 305-billion-parameter system under the permissive MIT license. That moves it from API-only access toward self-hosting—but the freedom comes with a 168-gigabyte checkpoint and a serious hardware bill. The repository includes 48 Safetensors files, a tokenizer, a prompt-encoding reference, and minimal PyTorch inference code. It also provides deployment instructions for vLLM and SGLang, so infrastructure teams can inspect, run, and adapt the model rather than route every image-and-text request through DeepSeek. The timing is notable. API access opened on August 21, and the weights and reference code followed ten days later. The license allows broad modification and reuse, but the model is still aimed mainly at research groups, inference providers, and developers with multi-GPU systems. Hugging Face listed no hosted inference provider for the checkpoint at publication, leaving the API as the easier option for most individual developers. That API supports OpenAI- and Anthropic-compatible interfaces, with up to 600 images in a request. But each image is capped at 384 processing tokens, so that limit describes input capacity—not visual accuracy. DeepSeek reports wins over Claude Opus 4.8 on two listed tests, but one win and six losses across seven text-agent comparisons. The key constraint is straightforward: open weights are available, but operating them independently still requires substantial multi-GPU infrastructure.

Story brief

3 key points

DeepSeek has moved V4 Flash Vision Exp from hosted API toward self-hosting by releasing its 305B-parameter checkpoint, code, and serving guidance under MIT. The 168GB package broadens inspection and deployment options, but effectively targets multi-GPU operators rather than typical developers; no hosted Hugging Face inference provider was listed at publication. API access remains easier, with OpenAI- and...

  1. 01

    API access began August 21; open weights and reference code followed 10 days later.

  2. 02

    The repository contains 48 Safetensors files, a tokenizer, prompt-encoding reference, and PyTorch inference code.

  3. 03

    DeepSeek caps each image at 384 processing tokens, despite allowing up to 600 images per request.

DeepSeek has published downloadable weights for its experimental V4 Flash vision model under the MIT license, moving the 305-billion-parameter system from API-only access toward self-hosted deployment. The shift gives developers more control over a model that accepts images, but the 168GB checkpoint sets a high infrastructure threshold for anyone seeking to run it locally.

What the download actually contains

DeepSeek-V4-Flash-Vision-Exp is based on the V4 Flash architecture, with visual modules added and further training for image understanding. The repository packages its weights in 48 Safetensors files, alongside a tokenizer, prompt-encoding reference and a minimal PyTorch inference implementation.

That combination is more than a file drop. The release also includes instructions for serving the checkpoint with vLLM and SGLang, two systems used to operate models as services. Teams can therefore inspect, deploy and adapt the supplied checkpoint rather than send every image-and-text request through DeepSeek’s hosted product.

Access is open; operation is not effortless

A permissive license permits broad use, modification and incorporation of the checkpoint into other products. Yet a 168GB download is not a consumer-device deployment: the model is positioned for research groups, inference providers and developers with multi-GPU systems, while the hosted API will remain the simpler path for most individual developers.

  • Self-hosting can put the model’s execution and adaptation under an operator’s own infrastructure.
  • The release supplies serving guidance, but Hugging Face listed no hosted inference provider for this checkpoint at publication.
  • The API remains an alternative for applications that do not have the hardware needed to operate the downloaded model.

The vision interface has defined limits

The API accepts mixed image-and-text prompts through OpenAI-compatible Chat Completions and Responses interfaces, as well as an Anthropic-compatible Messages endpoint. Developers can provide JPEG, PNG, GIF or WebP images as base64 data, public URLs or references through DeepSeek’s Files API.

DeepSeek caps processing at 384 tokens per image after resizing, while a request can contain up to 600 images subject to file-size and request limits. Those numbers describe the API’s input boundaries, not a demonstrated measure of visual accuracy or reliability on a particular task.

A benchmark claim with an aging comparison

DeepSeek’s model card reports scores of 27.3 on Agents’ Last Exam and 35.0 pass-at-five on ZeroBench, versus 25.7 and 34.0 for Claude Opus 4.8. The same table places V4 Flash Vision behind Opus 4.8 on ApexBench and Chartography, and across seven text-agent tests DeepSeek reported one win and six losses.

Those are company-reported results, run with DeepSeek Harness’s minimal configuration, maximum reasoning effort, temperature 1.0 and top-p 0.95. They also compare the new release with Claude Opus 4.8, not Anthropic’s newer Opus 5, which Anthropic released on July 24 as an improvement over Opus 4.8.

Editorial analysis

Our Read

This is a distribution change more than a new vision launch: DeepSeek’s image-capable endpoint had already been available for 10 days, while the checkpoint now lets organizations move inference and modification onto their own infrastructure. The important test is whether serving providers add hosted access or whether teams with multi-GPU capacity begin adapting the model themselves. That divide will determine whether the MIT license creates broad practical access or chiefly more optionality for research groups and inference operators. DeepSeek’s earlier 384-token image cap also remains a useful constraint to watch in real multimodal workloads.

Related coverageDeepSeek Gives V4-Flash a Separate Vision API With a 384-Token Image Cap

Related coverage

Continue reading