DeepSeek V3 0324: Same Architecture, Better Post-Training

A dated snapshot like DeepSeek V3 0324 is easy to dismiss as a minor revision. It was not. With no change to the architecture and no change to the parameter count, DeepSeek moved AIME from 39.6 to 59.4 and GPQA from 59.1 to 68.4. It also changed the licence and fixed function calling. If you are deciding between the V3 checkpoints, this is the one that changed most per line of release notes.

It is also the release where a sampling quirk became important to understand, because people moving between DeepSeek's API and self-hosted copies of these weights get noticeably different behaviour from the same temperature setting.

What a dated snapshot actually is

DeepSeek announced the refresh on 25 March 2025; the identifier carries 0324. The weights are new but the network is the same shape as the original V3: a mixture-of-experts model with 671B total parameters and 37B active, listed as 685B on Hugging Face once the multi-token prediction module is included.

DeepSeek's announcement said API usage remained unchanged, which in practice meant its deepseek-chat alias moved to the new weights in place. On aggregators the snapshot keeps its own identifier, deepseek/deepseek-chat-v3-0324, so you can call it deliberately. That distinction is the whole point of a dated name: the alias follows the vendor's latest choice, and the snapshot stays put.

What changed, in numbers

The model card compares the two checkpoints directly:

• MMLU-Pro — 75.9 to 81.2 (+5.3)
• GPQA — 59.1 to 68.4 (+9.3)
• AIME — 39.6 to 59.4 (+19.8)
• LiveCodeBench — 39.2 to 49.2 (+10.0)

Beyond the benchmarks, the card lists better executability for front-end code and "more aesthetically pleasing" web pages and game front-ends, Chinese writing aligned with the R1 style, more detailed output for Chinese search and report analysis, and higher function-calling accuracy that fixes problems in earlier V3 versions.

The size of the AIME jump against an unchanged architecture is a useful reminder of how much capability lives in post-training rather than in parameter count. Pretraining vs post-training explains where each kind of behaviour comes from.

The licence moved to MIT

The original V3 shipped its weights under a custom DeepSeek Model Agreement. With 0324, DeepSeek released the models under the MIT licence, "just like DeepSeek-R1", covering both the repository and the weights.

For anyone self-hosting, redistributing a fine-tune or serving inference to customers, that is a real simplification. MIT has no use-based restrictions to audit and no custom terms to send to a lawyer. MIT-licensed models compared covers what that permits in practice.

The temperature trap

This is the part that costs people a day of confused debugging. DeepSeek's model card documents a mapping between the temperature you send to its API and the temperature the model actually samples at:

• For API temperatures from 0 to 1, the model temperature is the API value multiplied by 0.3.
• For API temperatures above 1 up to 2, the model temperature is the API value minus 0.7.

So an API temperature of 1.0, the default many SDKs send, becomes a model temperature of 0.3 on DeepSeek's own service. DeepSeek's web and app use 0.3 as well.

Now take the same weights and serve them yourself with a standard inference server. Send temperature 1.0 and the model samples at 1.0, more than three times hotter than on DeepSeek's API. Output becomes looser, code becomes less reliable, and it is easy to conclude that the self-hosted copy is broken or badly quantised when the real difference is a setting.

The fix is simple once you know: when self-hosting, set temperature around 0.3 to match DeepSeek's defaults. When calling through a third-party provider, do not assume either way. Send the same prompt at a few temperatures and compare variance against DeepSeek's own endpoint. Temperature, top-p and sampling explains what the knob actually changes.

Where it sits in the line today

V3 0324 is listed on OpenRouter at $0.29 per million input tokens and $1.14 per million output tokens at the time of writing, with cache reads at $0.11. Check current rates before you rely on them.

Its successor changes the picture for agent work. DeepSeek's V3.1 model card puts V3 0324 at 45.4 on SWE-bench Verified in agent mode against 66.0 for V3.1 in non-thinking mode, and at 13.3 on Terminal-bench against 31.3. Those are large gaps on exactly the tasks coding agents perform.

That gives a fairly clean decision rule:

1. Moving off the original V3 with a pipeline you want to change as little as possible: 0324 is the safe step. Same architecture, better scores, simpler licence.
2. Building anything agentic or tool-heavy: skip ahead to DeepSeek V3.1 or its Terminus update. The agent benchmarks are not close.
3. Pinned for reproducibility: keep 0324 and call it by its dated name, never through an alias.

Whatever you choose, re-check your temperature setting at the same time as the model id. Changing both together and testing once is how sampling problems get mistaken for model problems.