DeepSeek re-releases improved V3 model under MIT license

DeepSeek just rolled out an updated version of DeepSeek-V3, its flagship 671B open-source large language model that only activates around 37B parameters per token. It handles general text tasks, writes code, and solves math problems.

The headline shift is the license. DeepSeek dropped its custom terms and re-released the model under the permissive MIT license. Developers and companies can now modify, redistribute, and commercially deploy the weights with virtually zero legal strings attached.

Coding bumps and silent upgrades

The update was pushed quietly without an official announcement. Developer Simon Willison was first to spot the changes, noting that the repository's readme file was left empty.

According to community benchmarks highlighted by VentureBeat, the refreshed release scores around 60% on Python and Bash code generation tests. That represents a bump of several percentage points over the original December release.

Much of that improvement comes from fine-tuning. DeepSeek engineers used prompt responses generated by their reasoning flagship, DeepSeek-R1, to polish V3's output quality. That said, it still trails R1 and reasoning models like Qwen-32B on complex problem solving.

Running a 671B beast on a Mac Studio

Because DeepSeek-V3 only activates a fraction of its total parameters during generation, it is far lighter on compute and inference costs than traditional dense models.

Apple machine learning researcher Awni Hannun tested the new weights locally on a maxed-out Mac Studio priced at $9,499. By applying 4-bit quantization, he got the model running smoothly at roughly 20 tokens per second.

While 4-bit quantization trades away a tiny sliver of accuracy, it proves that a frontier-class model can run on local workstation hardware instead of an enterprise cloud cluster.

Why it matters

We think the licensing pivot is vastly more important than the minor benchmark gains. Switching to MIT removes the lingering compliance headaches that often keep enterprise teams from using open weights in production.

If you need a cost-effective, commercially clean foundational LLM that you can self-host or fine-tune, the updated DeepSeek-V3 is now one of the easiest recommendations on the market. Just keep in mind that if your workload strictly demands deep step-by-step reasoning, models like R1 still hold the upper hand.