DeepSeek releases MIT-licensed DeepSeek-V3-0324, the latest version of their enormous DeepSeek v3 model; the previous DeepSeek v3 version had a custom license
deepseek-ai/DeepSeek-V3-0324. Chinese AI lab DeepSeek just released the latest version of their enormous DeepSeek v3 model … X: @awnihannun , @simonw , @simonw , @iterintellectus , @levie , and @datachaz X: Awni Hannun / @awnihannun : The new Deep Seek V3 0324 in 4-bit runs at > 20 toks/sec on a 512GB M3 Ultra with mlx-lm! [video] Simon Willison / @simonw : Notes on today's DeepSeek v3 0324 model - a 641 GB MIT licensed monster, but you can run it on a ~$10,000 consumer level 512GB M3 Mac Studio if you use the 352 GB quantized version via MLX https://simonwillison.net/... Simon Willison / @simonw : Updated my post with more notes about ways to run it - it's live on @OpenRouter right now so you can chat with the new model at https://openrouter.ai/... Vittorio / @iterintellectus : deepseek, out of nowhere, dropping a new model ~700GB, mit license. incredible [image] Aaron Levie / @levie : It never stops being amazing that you can fit world's intelligence on a usb thumb drive that costs less than $90 Charly Wargnier / @datachaz : Seems like @OpenAI is getting DeepSeek'd... again! @Deepseek_ai just dropped a major upgrade to DeepSeek V3 - and it's way more powerful than the minor update we expected! 🤯 link below [image]
Context & Ripple Effects
DeepSeek-V3-0324 changes the access terms for the V3 line: the original 671B-parameter V3 release was described as open-source, but this update replaces its custom license with the widely understood MIT license. That makes the licensing shift as consequential as the model refresh for developers evaluating reuse and deployment.
The release also illustrates the tension between very large open-weight models and practical local inference: the coverage points to a quantized version running on a high-memory consumer Mac Studio. Subsequent V3 coverage continued to emphasize efficiency techniques and lower tool pricing, reinforcing that deployment economics—not benchmark claims alone—are central to the product arc.
First-order effects
- Developers and companies gain clearer permission to modify, redistribute, and embed DeepSeek-V3-0324 than under the prior custom V3 license, reducing legal-review friction for adoption.
- The model remains operationally demanding despite local quantization options; users must still weigh hardware memory requirements and quantization trade-offs before self-hosting it.
Second-order effects
- Competing model providers face greater pressure to pair capable weights with permissive, legible terms, since restrictive or bespoke licenses become a more visible adoption disadvantage.
- Hardware and inference software ecosystems can benefit when a large MIT-licensed model is feasible on high-memory local systems: users have a stronger incentive to optimize runtimes, quantization, and deployment stacks around it.
Third-order effects
- If large-model releases increasingly combine permissive licensing with workable local inference, model buyers gain leverage to self-host, customize, or switch providers rather than depend solely on hosted APIs.
- The limiting factor may shift from formal access to infrastructure access: permissive weights broaden distribution, but memory capacity, runtime efficiency, and chip availability still determine who can operate them economically.
The trend: This is one data point in the shift from open weights as a research signal toward permissively licensed models as deployable infrastructure, where licensing clarity and inference efficiency jointly shape adoption.