DeepSeek drops V3-0324 with MIT license and a 19.8-point math leap

The newest model from DeepSeek isn’t a flashy launch. It’s a silent update that speaks volumes. DeepSeek-V3-0324, an upgraded version of the base model powering the R1 reasoning system, just appeared on Hugging Face with zero fanfare but one massive change: an MIT license. That’s a shift from the previous V3’s custom license, and it instantly makes the model more attractive for commercial use and tinkering.

The numbers are the real story. On the AIME benchmark, a proxy for math capabilities, the model vaulted from 39.6 to 59.4—a 19.8-point jump. That’s a staggering improvement that puts it in a different class for mathematical reasoning. It also posted strong gains on GPQA (up 9.3 points to 68.4) and LiveCodeBench (up 10 points to 49.2), indicating its coding chops now land between Claude-Sonnet-3.7 and GPT-4.5. The DeepSeek team hasn’t published a technical report yet, but the architecture is identical to the original V3. That points to a combination of improved post-training data and potentially some continual pretraining on better-curated sources.

The practical upgrades are refreshingly specific. Front-end web development gets a boost with code that actually executes and produces more visually appealing pages. Chinese writing sees enhanced style that aligns with the R1 model’s output, particularly for longer pieces. Function calling accuracy has been explicitly addressed, fixing bugs from the previous V3 release.

You won’t need exotic hardware to take it for a spin. Hugging Face’s Inference Providers offer access through Fireworks, Hyperbolic, and Novita. TGI and SGLang both support running it on H100 nodes, and the Unsloth and Llama.cpp teams are working on dynamic quants to drop VRAM requirements significantly. For a model that can now trade blows with GPT-4.5 while carrying a permissive license, the quiet release strategy feels deliberate—letting the benchmarks do the talking while competitors spend millions on launch events.