# DeepSeek Releases 305B Multimodal Model with Major Agent Benchmark Gains

> DeepSeek is back with DeepSeek-V4-Flash-Vision-Exp, a 305-billion-parameter multimodal model released under a permissive MIT license. The headline is efficiency: only about 13 billion of its parameters are active at once, a sparse mixture-of-experts design that keeps inference cheap relative to the raw parameter count. The benchmark movement is what will get attention. On ApexBench, a measure of agent capability, Pass@1 jumped from 26.2 to 36.5, with smaller gains on the punishing Agents' Last Exam. Numbers like these are easy to over-read, but a ten-point leap on an agent benchmark from an openly licensed model is the kind of thing that reshapes what small teams can build without paying frontier-lab API prices. The broader story is that the open-weight frontier keeps being pushed by Chinese labs, and doing it under MIT licensing rather than the more restrictive terms some Western releases carry. For anyone tracking where usable, self-hostable AI is heading, an experimental release like this is a better signal than another closed model demo.

_Section: [Daily AI Updates](https://www.wortins.com/daily-ai) · Source: AI Weekly · Published Tuesday, September 1, 2026_

## Wortins' read

DeepSeek is back with DeepSeek-V4-Flash-Vision-Exp, a 305-billion-parameter multimodal model released under a permissive MIT license. The headline is efficiency: only about 13 billion of its parameters are active at once, a sparse mixture-of-experts design that keeps inference cheap relative to the raw parameter count. The benchmark movement is what will get attention. On ApexBench, a measure of agent capability, Pass@1 jumped from 26.2 to 36.5, with smaller gains on the punishing Agents' Last Exam. Numbers like these are easy to over-read, but a ten-point leap on an agent benchmark from an openly licensed model is the kind of thing that reshapes what small teams can build without paying frontier-lab API prices. The broader story is that the open-weight frontier keeps being pushed by Chinese labs, and doing it under MIT licensing rather than the more restrictive terms some Western releases carry. For anyone tracking where usable, self-hostable AI is heading, an experimental release like this is a better signal than another closed model demo.

## Source

[Read the full story at AI Weekly](https://aiweekly.co/ai-news-today)

## Related coverage

- [Chai Discovery Raises $400M Series C at $3.8B Valuation for AI Drug Discovery](https://www.wortins.com/story/chai-discovery-raises-400m-series-c-at-3-8b-valuation-for-ai-e854aad3) — [FierceBiotech](https://www.fiercebiotech.com/biotech/chai-brews-400m-series-c-fuel-ai-used-lilly-novartis-and-pfizer)
- [Okta Launches Agent SSO for Enterprise AI Agent Identity Management](https://www.wortins.com/story/okta-launches-agent-sso-for-enterprise-ai-agent-identity-man-842af454) — [Okta](https://www.okta.com/newsroom/press-releases/okta-brings-first-class-identity-to-ai-agents-with-agent-sso/)
- [Microsoft Releases Phi-4 Silica for Windows AI+ PCs with Independent Updates](https://www.wortins.com/story/microsoft-releases-phi-4-silica-for-windows-ai-pcs-with-inde-a9b8f1d0) — [Microsoft](https://blog.logrocket.com/ai-dev-tool-power-rankings/)
- [OpenAI Pauses Reinforcement Learning Training Over Cybersecurity Risks](https://www.wortins.com/story/openai-pauses-reinforcement-learning-training-over-cybersecu-fe50160c) — [OpenAI](https://helpnetsecurity.com)
- [Yellow.ai Announces SPAC Merger with Bluerock for $550M Public Offering](https://www.wortins.com/story/yellow-ai-announces-spac-merger-with-bluerock-for-550m-publi-a82f4e54) — [Business Standard](https://www.business-standard.com/content/press-releases-ani/yellow-ai-a-global-leader-in-enterprise-agentic-ai-to-go-public-via-550-million-merger-with-bluerock-acquisition-corp-126080301360_1.html)
- [EU Designates ChatGPT as Major Online Service Under Digital Services Act](https://www.wortins.com/story/eu-designates-chatgpt-as-major-online-service-under-digital--47b609ba) — [Greek City Times](https://greekcitytimes.com/2026/09/01/breaking-eu-designates-chatgpt-under-digital-services-act-as-reddit-and-roblox-named-very-large-online-platforms)

---

_Curated and written by [Wortins](https://www.wortins.com) — The daily AI briefing. Every story links to its original source; the "Wortins read" on each is our own original analysis. [About Wortins & our editorial approach](https://www.wortins.com/about)._