DeepSeek V3からV3.2へ：アーキテクチャ、スパースアテンション、RLアップデート

From DeepSeek V3 to V3.2: Architecture, Sparse Attention, and RL Updates

Understanding How DeepSeek's Flagship Open-Weight Models Evolved

Sebastian Raschka, PhDDec 03, 20252541328ShareLast updated: January 1st, 2026

Similar to DeepSeek V3, the team released their new flagship model over a major US holiday weekend. Given DeepSeek V3.2’s really good performance (on GPT-5 and Gemini 3.0 Pro) level, and the fact that it’s also available as an open-weight model, it’s definitely worth a closer look.

Figure 1: Benchmark comparison between DeepSeek V3.2 and proprietary flagship models. This is an annotated figure from the DeepSeek V3.2 report.

I covered the predecessor, DeepSeek V3, at the very beginning of my The Big LLM Architecture Comparison article, which I kept extending over the months as new architectures got released. Originally, as I just got back from Thanksgiving holidays with my family, I planned to “just” extend the article with this new DeepSeek V3.2 release by adding another section, but I then realized that there’s just too much interesting information to cover, so I decided to make this a longer, standalone article.

There’s a lot of interesting ground to cover and a lot to learn from their technical reports, so let’s get started!

1. The DeepSeek Release Timeline

While DeepSeek V3 wasn’t popular immediately upon release in December 2024, the DeepSeek R1 reasoning model (based on the identical architecture, using DeepSeek V3 as a base model) helped DeepSeek become one of the most popular open-weight models and a legit alternative to proprietary models such as the ones by OpenAI, Google, xAI, and Anthropic.

Figure 2: DeepSeek V3/R1 architecture from December 2024. We will revisit and discuss architectural details in a later section.

So, what’s new since V3/R1? I am sure that the DeepSeek team has been super busy this year. However, there hasn’t been a major release in the last 10-11 months since DeepSeek R1.

Personally, I think it’s reasonable to go ~1 year for a major LLM release since it’s A LOT of work. However, I saw on various social media platforms that people were pronouncing the team “dead” (as a one-hit wonder).

I am sure the DeepSeek team has also been busy navigating the switch from NVIDIA to Huawei chips. By the way, I am not affiliated with them or have spoken with them; everything here is based on public information. As far as I know, they are back to using NVIDIA chips.

Finally, it’s also not that they haven’t released anything. There have been a couple of smaller releases that trickled in this year, for instance, DeepSeek V3.1 and V3.2-Exp.

Figure 3: DeepSeek releases since last year. The main models are shown in red.

As I predicted back in September, the DeepSeek V3.2-Exp release was intended to get the ecosystem and inference infrastructure ready to host the just-released V3.2 model.

V3.2-Exp and V3.2 use a non-standard sparse attention variant that requires custom code, but more on this mechanism later. (I was tempted to cover it in my previous Beyond Standard LLMs article, but Kimi Linear was released around then, which I prioritized for this article section on new attention variants.)

1. Hybrid Versus Dedicated Reasoning Models

Before discussing further model details, it might be worthwhile to discuss the overall model types. Originally, DeepSeek V3 was released as a base model, and DeepSeek R1 added additional post-training to develop a dedicated reasoning model. This procedure is summarized in the figure below.

Figure 4: Overview of the DeepSeek R1 training pipeline. This figure is from my more detailed Understanding Reasoning LLMs article.

You can read more about the training pipeline in the figure above in my Understanding Reasoning LLMs article.

What’s worthwhile noting here is that DeepSeek V3 is a base model, and DeepSeek R1 is a dedicated reasoning model.

In parallel with DeepSeek, other teams have also released many really strong open-weight reasoning models. One of the strongest open-weight models this year was Qwen3. Originally, it was released as a hybrid reasoning model, which means that users were able to toggle between reasoning and non-reasoning modes within the same model. (In the case of Qwen3, this toggling was enabled via the tokenizer by adding/omitting <think></think>

Since then, LLM teams have released (and in some cases gone back and forth between) both dedicated reasoning models and Instruct/Reasoning hybrid models, as shown in the timeline below.

Figure 5: The timeline of some of the reasoning and hybrid models released this year.

For instance, Qwen3 started out as a hybrid model, but the Qwen team then later released separate instruct and reasoning models as they were easier to develop and yielded better performance in each respective use case.

Some models like OpenAI’s gpt-oss only come in a hybrid variant where users can choose the reasoning effort via a system prompt (I suspect this is handled similarly in GPT-5 and GPT-5.1).

And in the case of DeepSeek, it looks like they moved in the opposite direction from a dedicated reasoning model (R1) to a hybrid model (V3.1 and V3.2). However, I suspect that R1 was mainly a research project to develop reasoning methods and the best reasoning model at the time. The V3.2 release may be more about developing the best overall model for different use cases. (Here, R1 was more like a testbed or prototype model.)

And I also suspect that, while the DeepSeek team developed V3.1 and V3.2 with reasoning capabilities, they might still be working on a dedicated R2 model.

1. From DeepSeek V3 to V3.1

Before discussing the new DeepSeek V3.2 release in more detail, I thought it would be helpful to start with an overview of the main changes going from V3 to V3.1.

3.1 DeepSeek V3 Overview and Multi-Head Latent Attention (MLA)

I already discussed DeepSeek V3 and R1 in great detail in several other articles. To summarize the main points, DeepSeek V3 is a base model that uses two noteworthy architecture aspects: Mixture-of-Experts (MoE) and Multi-Head Latent Attention (MLA).

I think you are probably well familiar with MoE at this point, so I am skipping the introduction here. However, if you want to read more, I recommend the short overview in my The Big Architecture Comparison article for more context.

The other noteworthy highlight is the use of MLA. MLA, which is used in DeepSeek V2, V3, and R1, offers a memory-saving strategy that pairs particularly well with KV caching. The idea in MLA is that it compresses the key and value tensors into a lower-dimensional space before storing them in the KV cache.

At inference time, these compressed tensors are projected back to their original size before being used, as shown in the figure below. This adds an extra matrix multiplication but reduces memory usage.

(As a side note, the queries are also compressed, but only during training, not inference.)

Figure 6: Multi-Head Latent Attention (MLA) in DeepSeek V3/R1. (The compressed space of the query vector is not shown for simplicity.)

The figure above illustrates the main idea behind MLA, where the keys and values are first projected into a latent vector, which can then be stored in the KV cache to reduce memory requirements. This requires a later up-projection back into the original key-value space, but overall it improves efficiency (as an analogy, you can think of the down- and up-projections in LoRA).

Note that the query is also projected into a separate compressed space, similar to what’s shown for the keys and values. However, I omitted it in the figure above for simplicity.

By the way, as mentioned earlier, MLA is not new in DeepSeek V3, as its DeepSeek V2 predecessor also used (and even introduced) it.

3.2 DeepSeek R1 Overview and Reinforcement Learning with Verifiable Rewards (RLVR)

DeepSeek R1 uses the same architecture as DeepSeek V3 above. The difference is the training recipe. I.e., using DeepSeek V3 as the base model, DeepSeek R1 was focused on the Reinforcement Learning with Verifiable Rewards (RLVR) method to improve the reasoning capabilities of the model.

The core idea in RLVR is to have the model learn from responses that can be verified symbolically or programmatically, such as math and code (but this can, of course, also be extended beyond these two domains).

Figure 7: An example of a verifiable task.

The GRPO algorithm, which is short for Group Relative Policy Optimization, is essentially a simpler variant of the Proximal Policy Optimization (PPO) algorithm that is popular in Reinforcement Learning with Human Feedback (RLHF), which is used for LLM alignment.

Figure 8: Comparison of reinforcement learning setups in LLM training. Traditional RLHF with PPO uses both a reward model (trained on human preferences) and a critic (value model) to guide learning. GRPO eliminates the critic model. RLVR with GRPO goes a step further by removing the reward model, relying instead on verifiable rewards from symbolic tools such as calculators or compilers.

I covered the RLVR training with their GRPO algorithm in more detail (including the math behind it) in my The State of Reinforcement Learning for LLM Reasoning if you are interested in additional information.

3.3 DeepSeek R1-0528 Version Upgrade

As the DeepSeek team stated themselves, DeepSeek R1-0528 is basically a “minor version upgrade.”

The architecture remains the same as in DeepSeek V3/R1, and the improvements are on the training side to bring it up to par with OpenAI o3 and Gemini 2.5 Pro at the time.

Unfortunately, the DeepSeek team didn’t release any specific information describing how this was achieved; however, they stated that it partly comes from optimizations in their post-training pipeline. Also, based on what’s been shared, I think it’s likely that the hosted version of the model uses more computational resources at inference time (longer reas