DeepSeek V3 (deepseek-chat): The December 2024 Original

"deepseek-chat" is one of the most confusing model names in circulation, because it has meant different models in different places. On OpenRouter, deepseek/deepseek-chat is the original DeepSeek V3, released on 26 December 2024. On DeepSeek's own API, deepseek-chat was a moving alias that was upgraded in place several times and ended up pointing at the non-thinking mode of DeepSeek V4 Flash, with discontinuation announced for 24 July 2026.

This guide is about the first of those: the December 2024 model itself. It matters historically, it still turns up in pinned pipelines, and understanding it explains most of what came after in the DeepSeek line.

Which model this name points to

If you send requests to deepseek/deepseek-chat on OpenRouter, the catalogue describes the model as DeepSeek V3 with a release date of 26 December 2024. That is the original checkpoint, not the March 2025 refresh and not V3.1.

If you have old code that called DeepSeek's API directly with model: "deepseek-chat", the model behind it changed over time without the name changing: V3, then V3-0324, then the non-thinking mode of V3.1 and its successors, then V4 Flash. In April 2026 DeepSeek announced that the legacy deepseek-chat and deepseek-reasoner names would be discontinued on 24 July 2026, so treat any code still sending them as broken until you have confirmed otherwise.

This is the textbook case for pinning model versions. An alias that upgrades silently is convenient until an evaluation you ran last quarter no longer describes the model you are calling.

The architecture that set the template

DeepSeek V3 is a mixture-of-experts model with 671B total parameters and 37B activated per token. The Hugging Face repository totals 685B because it includes a 14B multi-token prediction module. The pieces that made it notable:

• Multi-head latent attention (MLA), which compresses the KV cache. Multi-head latent attention explained covers why that matters for long context.
• DeepSeekMoE with an auxiliary-loss-free load-balancing strategy, avoiding the quality penalty that a balancing loss term can impose. Load balancing in MoE has the background.
• A multi-token prediction objective during training.
• FP8 mixed-precision training, which DeepSeek described as the first validation of FP8 training at that scale.

Pre-training used 14.8 trillion tokens. DeepSeek reported 2.788 million H800 GPU hours for the full training run, of which 2.664 million were pre-training. That compute figure became one of the most quoted numbers of early 2025, because it was far lower than what people assumed a model of this capability required.

The model card gives a 128K-token context window. DeepSeek's launch note quoted generation at 60 tokens per second on its own service, three times faster than V2.

What it scored in 2024

DeepSeek's model card lists these results for the chat model:

• MMLU — 88.5; MMLU-Redux — 89.1; MMLU-Pro — 75.9
• GPQA-Diamond — 59.1
• MATH-500 — 90.2; AIME 2024 — 39.2
• HumanEval-Mul — 82.6
• SWE-bench Verified — 42.0

Those were strong numbers for an open model at the end of 2024. They look modest now. The same lab's DeepSeek V4 Pro reports 80.6 on SWE-bench Verified, nearly double the original V3's 42.0. Even within the V3 line, the March 2025 refresh moved GPQA from 59.1 to 68.4 without changing the architecture.

The licence was not MIT at first

This catches people who assume every DeepSeek model is MIT-licensed. For the original V3, the code repository was MIT, but the weights were released under a separate DeepSeek Model Agreement. That agreement permits commercial use, but it is a custom licence with its own terms, and it is worth reading before you build on the original checkpoint.

DeepSeek moved the V3 line to MIT for both code and weights with the March 2025 refresh. If licence simplicity matters to you, that alone is a reason to prefer DeepSeek V3 0324 over the original.

Why anyone still calls it

The honest answer is that there are few reasons to start new work on the original V3. It is listed on OpenRouter at $0.32 per million input tokens and $0.89 per million output tokens at the time of writing, with a 16,384-token output cap. Later models in the same family are priced similarly and score better.

The legitimate reasons are about stability rather than capability:

• Reproducing an old result. If a paper, an internal evaluation or a customer-facing baseline was produced on V3, you need V3 to reproduce it.
• A tuned pipeline. Prompts, parsers and output post-processing tuned to one model's quirks can break on a stronger model. Sometimes the cheapest option this quarter is to leave it alone.
• Comparisons. It is a useful fixed reference point when measuring how far later models have moved on your own tasks.

Moving off it

If you are still on the original, pick the destination by what you need:

1. Same architecture, better reasoning and an MIT licence: move to V3 0324. It is the lowest-risk step.
2. Optional reasoning and much stronger agent scores: move to V3.1, which merges chat and reasoning modes in one model.
3. Current generation: move to DeepSeek V4 Flash, which offers a 1M-token context.

Whichever you choose, run your own regression set first. A stronger model on average is still a different model on your edge cases, and regression testing a model upgrade shows how to catch the cases that get worse.