Chinese AI labs in 2026: DeepSeek, Qwen and Kimi releases and direction

Three of China's best known AI labs, DeepSeek, Alibaba's Qwen team and Moonshot AI (Kimi), release large models on a fast cadence, mostly as open weights, with a stated focus on efficiency and long context. This post sets out what their own pages say, with dates, and flags every claim that comes from the labs themselves.

The facts below come from official lab pages, documentation, model cards and one Alibaba filing, as they stood on 19 October 2026. Benchmark scores and "state of the art" statements are the labs' own and have not been independently verified. This post does not rank the labs. For how they price their APIs and how Alibaba pays for its compute, see the companion post on how Chinese AI labs fund and price their models.

Key numbers at a glance

• 29 September 2025 (DeepSeek docs): DeepSeek-V3.2-Exp introduced DeepSeek Sparse Attention for long inputs.
• 24 April 2026 (DeepSeek docs): DeepSeek-V4 released in two sizes. V4-Flash has 284 billion total parameters with 13 billion active.
• 1.6 trillion total parameters, 49 billion active (DeepSeek-V4-Pro model card on Hugging Face, checked 19 October 2026): the stated size of V4-Pro, under the MIT license.
• 1 million tokens (DeepSeek docs, 24 April 2026): the context length, which DeepSeek says is now the default across its official services.
• 27 percent and 10 percent (DeepSeek-V4-Pro model card): the card's claim that at 1 million tokens of context the model needs 27 percent of the single-token inference compute and 10 percent of the memory cache that DeepSeek-V3.2 needs.
• 17 October 2026 (DeepSeek updates page): DeepSeek-V4.1-Flash-1017, the newest dated entry as of 19 October 2026.
• 29 April 2025 (Qwen3 repository): Qwen3 released, with dense models from 0.6 billion to 32 billion parameters and mixture of experts models of 30B-A3B and 235B-A22B, under Apache 2.0.
• 397 billion total, 17 billion active (Simon Willison's report of the Qwen3.5 launch, 17 February 2026): the first open Qwen3.5 model.
• 2.4 trillion total, 95 billion active (Qwen3.8-2.4T-A95B model card, cited as August 2026): the open model behind Qwen3.8-Max.
• 1 trillion total, 32 billion active (Kimi K2 Instruct model card): Kimi K2, released under a Modified MIT License.
• 2.8 trillion total, 104 billion active (Kimi-K3 model card on Hugging Face): Kimi K3, with 16 of 896 experts selected per token and a 1 million token context.

Terms you need

Parameters are the numbers a model learns during training. More parameters can mean more capacity, but not automatically better answers.

Mixture of experts (MoE) is a design where a model contains many specialist sub-networks, called experts, and only some are used for each token. That is why you see two numbers: total parameters and "active" parameters per token. A model with 1.6 trillion total and 49 billion active costs far less to run per token than a dense model of 1.6 trillion, though all 1.6 trillion still have to sit in memory somewhere.

Dense models use every parameter for every token. They are simpler to serve but cost more per token at the same total size.

Open weights means the trained model files can be downloaded and run by anyone, subject to the license. It is not the same as open source in the full sense, because training data and code may not be released.

Context length is how much text the model can consider at once. A million tokens is roughly several novels. Long context is expensive because the model keeps a memory of earlier text, called the KV cache, which grows with length.

FP4 and FP8 are number formats that use 4 and 8 bits. Smaller formats let hardware do more arithmetic per second and use less memory, with some loss of precision. MXFP4 and MXFP8 are variants that share a scale factor across small blocks of numbers.

A token is a small chunk of text, often part of a word. Model sizes, context windows and prices are all counted in tokens.

DeepSeek: efficiency as the stated theme

DeepSeek's documentation gives a dated trail, which makes it easy to follow.

On 20 January 2025 it released DeepSeek-R1, a reasoning model. The R1 model card on Hugging Face states that the repository and weights are under the MIT license and that the license permits derivative works, including "distillation for training other LLMs." The card claims performance "comparable to OpenAI-o1 across math, code, and reasoning tasks," which is DeepSeek's own claim, and lists six smaller dense models distilled from R1 at 1.5, 7, 8, 14, 32 and 70 billion parameters, built on Qwen2.5 and Llama bases. On the API, R1 is served under the name deepseek-reasoner.

On 29 September 2025 it released DeepSeek-V3.2-Exp, called an experimental model built on V3.1-Terminus. The new part was DeepSeek Sparse Attention, which the page says enables "faster, more efficient training and inference on long context" with little effect on output quality. The page says benchmarks showed parity with V3.1-Terminus and that weights were published on Hugging Face with a technical report and GPU kernels. DeepSeek cut API prices on the same day; the size of that cut is covered in the pricing companion post.

Sparse attention means the model looks closely at only the most relevant earlier tokens rather than all of them, which cuts the cost of long inputs.

On 24 April 2026 DeepSeek released V4 as a preview with two models. The V4-Pro model card on Hugging Face states "DeepSeek-V4-Pro with 1.6T parameters (49B activated)." V4-Flash has 284 billion total with 13 billion active. Both support 1 million tokens, and the release page says that is now the default across all official services. The release headline reads "Welcome to the era of cost-effective 1M context length." It describes combining token-wise compression with sparse attention to lower compute and memory costs. DeepSeek says V4-Pro is open-source state of the art on agentic coding benchmarks and trails only one named competing model on world knowledge. These are company claims.

A note on the size figure. The Hugging Face page for the later DeepSeek-V4-Pro-0813 checkpoint shows "1.7T params" in its automatically generated file metadata, which counts every stored tensor. Its text states no parameter count. The figure DeepSeek itself states, in the V4-Pro model card text, is 1.6 trillion total and 49 billion active, and that is the figure used across this series.

The API accepts both OpenAI-style and Anthropic-style request formats, so developers can switch by changing the model name. The old names deepseek-chat and deepseek-reasoner were scheduled for retirement on 24 July 2026.

The V4-Pro model card says the model uses mixed FP4 and FP8 precision, with mixture of experts parameters in FP4 and most other parameters in FP8, that it is released under the MIT license, and that a technical report is on arXiv. It also makes the efficiency claim in the Key numbers above, which is the card's own and unverified.

DeepSeek's updates page then lists a run of releases in 2026:

• 31 July 2026: V4-Flash public beta, re-post-trained, with large agent benchmark gains claimed.
• 13 August 2026: V4-Pro general release, with stronger agent capabilities, Responses API support and low, high and max thinking effort levels. The same entry introduced peak and off-peak pricing.
• 21 August 2026: V4-Flash-Vision-Exp, an experimental multimodal model.
• 10 September 2026: V4.1-Flash, described as the smallest model in the new architecture family, with native multimodal support.
• 17 October 2026: V4.1-Flash-1017, an upgrade of the Flash model, alongside the opening of DeepSeek's in-house search API.

Taken together, the pages describe a lab that cuts the cost of using a model through architecture, then cuts price. The DeepSeek pages cited here do not say which chips the V4 models were trained on, so this post makes no claim about that.

Qwen: a wide family under mostly open licenses

Alibaba's Qwen team publishes through GitHub, Hugging Face and its own site. The pattern across its releases is breadth: many sizes, dense and mixture of experts, text and vision, most of them downloadable.

Qwen3, April to August 2025. The Qwen3 repository states that Qwen3 was released on 29 April 2025, with dense models of 0.6, 1.7, 4, 8, 14 and 32 billion parameters and MoE models of 30B-A3B and 235B-A22B. The naming gives total parameters and, after the A, active parameters, so 235B-A22B has 235 billion total with 22 billion active. The repository says "All our open-weight models are licensed under Apache 2.0," a permissive license that allows commercial use with conditions. It lists support for more than 100 languages and dialects. Updated 2507 models followed between 21 July and 8 August 2025, with a 256K token context extendable to 1 million tokens. Original Qwen3 models can switch between thinking and non-thinking modes.

Qwen3.5, February 2026. Simon Willison reported on 17 February 2026 that Alibaba had released two Qwen3.5 models: an open weights model, Qwen3.5-397B-A17B, and a hosted model, Qwen3.5 Plus. He reported 397 billion total parameters with 17 billion active per forward pass, about 807 GB of weights on Hugging Face, a native 256K token context for the open model and a 1 million token window for the hosted version, with vision input on both. He quotes Qwen's Junyang Lin describing Plus as "a hosted API version of 397B." This is a secondary report; the license of that model is not stated in it.

Qwen3.8, August 2026. By August the Qwen organization page on Hugging Face listed a Qwen3.8 collection. Three model cards stand out:

• Qwen3.8-2.4T-A95B. A text-only model with 2.4 trillion total parameters and 95 billion active, 92 layers, and a 262,144 token native context extendable to about 1 million. Its layers alternate Gated DeltaNet blocks, a linear attention style that keeps a running summary instead of a full cache, with gated attention blocks, each followed by a mixture of experts block with 512 experts. Thinking mode cannot be switched off. The card says "Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B," and lists the license as "qwen3.8-max," a custom license rather than Apache 2.0.
• Qwen3.8-27B. A dense vision-language model of 27 billion parameters, under Apache 2.0, with the same context range.
• Qwen3.8-Flash-Next. About 125 billion parameters with 6 billion active per token, under a license named qwen-community-1.0. The card calls it "an experimental preview of the architecture that will underpin Qwen4" and says the official API model, Qwen3.8-Flash, is built on it. Its hybrid attention pairs Gated DeltaNet with what Qwen calls Qwen Sparse Attention.

Alibaba's own results release for the quarter ended 30 June 2026 confirms the flagship cadence: it says Qwen3.8-Max launched "within three months of its prior version" and that the company opened model weights of 2.4 trillion parameters.

The stated direction is evident in what these pages emphasize: a broad range of sizes, many languages, long context, and open weights, though the licenses are no longer uniformly Apache 2.0. Anyone choosing a Qwen model now has to check which of at least three license names applies.

Kimi: scale and energy efficiency

Moonshot's blog index lists these releases, with dates:

• Kimi K2, 11 July 2025.
• Kimi K2-Instruct-0905, 5 September 2025.
• Kimi K2 Thinking, 6 November 2025.
• Kimi K2.5, 27 January 2026.
• Kimi K2.6, 20 April 2026.
• Kimi K3, 16 July 2026.

The index states Moonshot's research goal: "Our research seeks the optimal conversion from energy into intelligence."

K2. Moonshot says K2 is a mixture of experts model with 1 trillion total and 32 billion active parameters, pre-trained on 15.5 trillion tokens. It says it developed an optimizer improvement called MuonClip, which controls unstable attention values, and reports "zero training spike." It states K2 Instruct reached 65.8 percent on a coding benchmark in a single attempt. Those are Moonshot's evaluations, and the comparison models were non-thinking models. The Kimi K2 Instruct model card on Hugging Face lists a 128K context and says "Both the code repository and the model weights are released under the Modified MIT License."

K3. The company's K3 blog post describes 2.8 trillion total parameters, 16 of 896 experts activated per token, a 1 million token context and native vision. The Kimi-K3 model card on Hugging Face gives 104 billion active parameters and adds that 2 of the experts are shared, meaning they run on every token. The blog says training uses MXFP4 weights with MXFP8 activations from the supervised fine-tuning stage, and it recommends deployment on "supernode configurations with 64 or more accelerators." The blog does not name the hardware, though the model card's footnotes say some evaluations ran on NVIDIA H20 GPUs.

The blog says the model's "overall performance still trails the most powerful proprietary models," naming two US models. That is an unusually direct statement from a lab about its own position. It also said full weights would be released by 27 July 2026. The Kimi-K3 repository on Hugging Face is live and says "Both the code repository and the model weights are released under the Kimi K3 License," a custom license whose terms are in the repository rather than summarized on the card.

What these labs have in common

Reading their pages side by side, four themes recur.

1. Efficiency before size. Sparse attention, linear attention hybrids, MoE routing, mixed low precision formats and token compression all aim to cut the cost per answer.
2. Long context as a default. DeepSeek V4, Kimi K3 and the hosted Qwen models all state a window of about 1 million tokens.
3. Open weights, with varied licenses. DeepSeek uses MIT, Kimi uses a Modified MIT License for K2 and a custom Kimi K3 License for K3, and Qwen spans Apache 2.0, a community license and a custom Max license.
4. Frequent dated releases. DeepSeek posted five dated entries between July and October 2026 alone, and Qwen shipped three Qwen3.8 model families in one month.

Why chips and policy show up in these stories

Efficiency is not only a cost goal. As the posts on China's domestic AI chips and US export controls explain, access to the newest chips is limited. A lab that needs less compute per answer is less exposed to chip supply. That is a reasonable reading of the incentives, not something these labs state; none of the pages cited here links release choices to chip access.

For the compute side outside China, see the explainers on Anthropic's compute commitments and OpenAI's compute commitments.

What is firm and what is uncertain

Firm, because the labs say so on their own pages or in a filing:

• Release dates, model sizes, context lengths and license names for DeepSeek V4, Qwen3, Qwen3.8 and Kimi K2 and K3.
• DeepSeek-V4-Pro's stated size of 1.6 trillion total and 49 billion active parameters.
• Each lab's stated focus on efficiency and long context.

Uncertain:

• All benchmark claims, which are self-reported and have not been independently verified.
• Which chips each lab used for training. None of the cited pages says.
• The exact terms of the Kimi K3 License and the qwen3.8-max license, which need a full reading before commercial use.
• Qwen3.5 details, which rest on a secondary report rather than a Qwen page.
• Training cost, usage and revenue for DeepSeek and Moonshot. No audited figures are public.

What to watch next

• DeepSeek's next dated entry on its updates page, which has moved roughly monthly in 2026.
• Qwen4, which the Qwen3.8-Flash-Next card names as the target of its architecture preview.
• Technical reports that describe training hardware, if any are published.
• Independent evaluations of these models by third-party labs.
• Whether domestic chips are named as deployment targets in future releases.

If you plan to run an open weights model inside your company, treat it like any other third-party component and record its source, version and license; Safeguard's free scan at https://gold.safeguard.sh/scan covers the open source packages around it.

Quick check

Sources

• DeepSeek API docs, updates page (accessed 19 October 2026)
• DeepSeek, DeepSeek-V4 preview release, 24 April 2026
• DeepSeek, DeepSeek-V3.2-Exp, 29 September 2025
• DeepSeek, DeepSeek-R1 release, 20 January 2025
• DeepSeek-R1 model card on Hugging Face
• DeepSeek-V4-Pro model card on Hugging Face
• DeepSeek-V4-Pro-0813 page on Hugging Face
• Qwen3 repository on GitHub
• Simon Willison, Qwen3.5 launch, 17 February 2026
• Qwen organization on Hugging Face (accessed 19 October 2026)
• Qwen3.8-2.4T-A95B model card on Hugging Face
• Qwen3.8-27B model card on Hugging Face
• Qwen3.8-Flash-Next model card on Hugging Face
• Alibaba Group, results for the quarter ended 30 June 2026 (Form 6-K, Exhibit 99.1), 20 August 2026
• Moonshot AI, Kimi blog index
• Moonshot AI, Kimi K2 blog
• Kimi K2 Instruct model card on Hugging Face
• Moonshot AI, Kimi K3 blog
• Kimi-K3 model card on Hugging Face

Last reviewed: 19 October 2026.