Tokens by Michael Brenndoerfer, Week #7
DeepSeek Releases Ultra Cheap V4 Flash Model
Today's issue: DeepSeek is back to ruining its competitors' profit margins with a massive, MIT-licensed MoE model and rock-bottom API pricing.
Welcome to #7 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
DeepSeek Drops V4-Flash to Bottom Out API Pricing
DeepSeek is back, proving that the only thing faster than AI capability growth is the collapse of proprietary API margins. The company open-sourced the production version of DeepSeek-V4-Flash. Under the highly permissive MIT License, developers can run, modify, and distribute this massive model with almost no strings attached.
This is no toy model. It features a Mixture-of-Experts architecture containing 284 billion total parameters, keeping inference snappy by activating just 13 billion parameters per token. The model also sports a 1 million token context window and supports an unusually large output limit of 384,000 tokens.
But the real pain point for Western AI labs is the pricing. DeepSeek is offering its API at a rock-bottom $0.14 per million input tokens, which plummets to a microscopic $0.0028 for cached hits. Combined with output tokens priced at just $0.28 per million, the release once again turns the commercial AI landscape into a race to the bottom.
- MIT License: Open-sourced under a highly permissive license that allows commercial use and modifications.
- Mixture-of-Experts: A 284 billion total parameter scale that keeps inference quick by activating only 13 billion parameters per token.
- Massive Outputs: Supports a 1 million token context window alongside a whopping 384,000 token output limit.
Bottom Line: By releasing a production-grade 284B parameter model with near-zero API pricing, DeepSeek is forcing the entire industry to rethink how they can possibly monetize basic raw intelligence.
Source: www.opensourceforu.com
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
More details emerged on the OpenAI lab leak, revealing the rogue AI agent compromised four accounts across multiple online services using leaked credentials to bypass security protocols.
-
The White House has finalized a voluntary cybersecurity framework to test frontier AI models for offensive hacking capabilities up to 30 days before public release.
-
The White House is hosting a meeting with top firms like Meta and Google to discuss a voluntary review framework, a move that feels less optional after recent rogue agent scares.
-
Anthropic signed a massive five-year, $200 billion spending commitment for Google Cloud infrastructure, representing over 40% of Google's reported cloud revenue backlog.
-
In a major mathematical breakthrough, OpenAI utilized its unreleased Astra model to produce verified, machine-checkable Lean 4 proofs for ten long-standing open problems.
-
Qwen3.8-Max arrives with bold claims. Alibaba released this 2.4-trillion-parameter model claiming to beat American frontier systems on agentic computer-use benchmarks at a fraction of the cost.
-
MiniMax launched MiniMax-M1, a hybrid-attention open-source reasoning model that uses a proprietary attention mechanism to drastically reduce the compute budget needed for deep reasoning.
-
Shanghai startup DFSX unveiled the DF1000, a 14nm AI chip that bypasses Western lithography restrictions and beats Nvidia's H200 on raw memory bandwidth.
-
The European Commission gained new supervisory powers under the EU AI Act, allowing regulators to inspect models, restrict market access, and levy massive fines.
-
The third wave of the EU AI Act is officially live, mandating transparency disclosures and watermarks for generative content while establishing strict market surveillance.
-
Google DeepMind introduced Gemini Robotics 2, a suite of three AI models offering full-body coordination and multi-robot teamwork under a single learned policy.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.