Chinese AI Models Compared: DeepSeek, Qwen, Kimi, GLM, MiMo, Doubao and Manus

Amna ManzoorLast Updated on November 5, 2026

13-14 Min Read TimeAdd as preferred on Google

Key Takeaways

As of November 2026, DeepSeek V4 and GLM-5.2 are the MIT-licensed starting points among Chinese AI models, and licensing and data location decide most evaluations.

• Start the shortlist with DeepSeek V4 or GLM-5.2, both MIT-licensed, and test it against your current provider.
• Kimi K3 requires a separate agreement with Moonshot AI for model-as-a-service providers whose revenue passes $20 million over any 12 months, and the 2.4T Qwen3.8 model has its own license.
• Doubao-Seed-2.0 has no open weights.
• Every Doubao prompt goes to ByteDance's cloud. Clear data location before testing it with real data.
• Self-hosting a large mixture-of-experts model means holding every weight in GPU memory. DeepSeek V4-Pro needs room for 1.6 trillion parameters even though only 49 billion run per token.
• The labs supply most benchmark figures themselves. A few hundred real requests from your own product are a better guide.
• Versions change quickly. DeepSeek's deepseek-flash endpoint has swapped its underlying model twice since September, and Xiaomi retired MiMo V2.5 about six months after public testing began.

Clear the license and data-location questions before any pilot touches production data.

Most engineering teams still picture DeepSeek R1 when someone mentions Chinese AI models. That picture is almost two years old. As of November 2026, this guide covers six Chinese labs whose models a team can test this week, and five of them release open weights for at least part of their line-up.

Two questions follow for a CTO or AI lead with a growing inference bill. Which two models belong on an evaluation shortlist, and what do the licenses on those models actually allow a company to do? This guide compares DeepSeek, Qwen, Kimi, GLM, Xiaomi MiMo and Doubao, plus the Manus agent, and then covers the license and data-location checks to clear before any of them sees production data.

Chinese AI models compared at a glance

DeepSeek V4, GLM-5.2 and Xiaomi MiMo-V2.6 carry the fewest license conditions among Chinese AI models in November 2026, because all of them ship under MIT. Kimi K3 and the largest Qwen3.8 model are open but use custom licenses.

Doubao is hosted only. Manus is an agent product that runs on top of models. The specifications below come from each maker's model cards and announcements, and most of the benchmark numbers quoted in this guide are the labs grading their own homework.

1. DeepSeek V4 and the R1 reasoning model

Specifications

The current open family is DeepSeek V4, first released as a preview on 24 April 2026. V4-Pro has 1.6 trillion parameters with 49 billion active per token, V4-Flash has 284 billion with 13 billion active, and both read up to one million tokens in a single request. Each uses a mixture-of-experts (MoE) design. In an MoE model, each token passes through only a few specialist sub-networks, which keeps compute per token low.

The official V4-Pro release on 13 August 2026 replaced the preview without changing its structure, and since 17 October 2026 the API's deepseek-flash endpoint has been serving DeepSeek-V4.1-Flash-1017. R1 is older. The reasoning model behind the 2025 headlines has 671 billion parameters with 37 billion of them active. Its context window is 128K tokens.

Best fit

This is the pick for teams that want to self-host or fine-tune a frontier-size model with as few legal conditions as possible.

Benefits

The V4 weights and R1 both ship under MIT, which lets a company use and modify them commercially with little beyond keeping the original copyright notice in place. Smaller budgets have an option too. DeepSeek's distilled R1 checkpoints range from 1.5B to 70B parameters and are built on Qwen2.5 and Llama3.

Caveats

Self-hosting V4-Pro is a serious infrastructure project. All 1.6 trillion weights have to sit in GPU memory even though only 49 billion run per token, and DeepSeek recommends a context window of at least 384K tokens for its maximum reasoning mode.

Versions also move fast. The deepseek-flash endpoint has already changed its underlying model twice since September. Pin a specific model version for any production test.

2. Qwen3.8 from Alibaba Cloud

Specifications

Alibaba published the weights for Qwen3.8, its current open family, on Hugging Face and ModelScope in August 2026. The two models in the family have little in common. Qwen3.8-27B is a dense vision-language model that understands images and video, with a native context of 262,144 tokens that YaRN scaling (a method for stretching a model's position encoding) extends to one million.

The other is Qwen3.8-2.4T-A95B, the Max-class model, with 2.4 trillion parameters and 95 billion active. It is text only and runs every request in thinking mode.

Best fit

Qwen fits teams that want one family covering both ends, a 27B model for their own hardware and a frontier-size one.

Benefits

Qwen3.8-27B ships under Apache 2.0, a permissive license that lets a company use it commercially. At 27 billion parameters, it's also the easiest model in this guide to run on your own hardware, and Alibaba's model card reports it beating Qwen3.7-Plus on agentic coding tests, for example 61.7 against 57.6 on SWE-bench Pro.

Caveats

The two open models don't share a license. The 2.4T model card lists a separate Qwen3.8-Max license, and until legal has read it, nobody should plan commercial use of the large model, however good the 27B results look. Alibaba's scores also come from its own test setups and include in-house benchmarks. For how the earlier generation stacked up, see our comparison of Alibaba's Qwen3 models with Claude Opus 4.

3. Kimi K3 from Moonshot AI

Specifications

Kimi K3 is a 2.8-trillion-parameter MoE model with 104 billion active parameters, a context window of 1,048,576 tokens and, according to the model card, text and image input. Moonshot AI calls it the world's first open 3T-class model. That's Moonshot's own label. K3 follows the Kimi k1.5 generation, covered in our earlier Kimi k1.5 review.

Best fit

K3 fits research and analysis work that puts long documents and images into a single prompt.

Benefits

K3 is the largest open model in this guide. Its one-million-token window can hold a full set of contracts or a mid-sized codebase in one request, which removes much of the chunking logic (splitting text into pieces and fetching the right ones) that a retrieval pipeline would otherwise need.

Caveats

The Kimi K3 License is custom. Its text exempts internal use, but a company that runs a model-as-a-service business with revenue above $20 million over any 12 months must sign a separate agreement with Moonshot AI before using K3 commercially. Products above 100 million monthly active users also have to display "Kimi K3" in their interface. Have counsel read it early.

4. GLM-5.2 from Zhipu AI

Specifications

GLM-5.2 is the most recent MIT-licensed GLM model from Zhipu AI (which now publishes as Z.ai), and Zhipu describes it as built for long-horizon tasks, meaning jobs that run through many steps before they finish. It builds on GLM-5 (744 billion parameters, 40 billion active) and adds a one-million-token context. Both ship under MIT. Its successor, GLM-5.3, released open weights in August 2026 under a custom GLM-5.3 License.

Best fit

It suits engineering teams building coding agents or multi-step automations on an MIT-licensed base.

Benefits

With about 40 billion active parameters, the GLM-5 design should cost less per token to serve than the larger MoE models here, such as Kimi K3 with its 104 billion. Zhipu's own results put GLM-5.2 at 81.0 on Terminal Bench 2.1 and 62.1 on SWE-bench Pro. It also has several thinking-effort levels. With them, a team can trade answer quality for latency request by request.

Caveats

Every GLM-5.2 score is self-reported. To its credit, Zhipu's own table shows Claude Opus 4.8 ahead on SWE-bench Pro, at 69.2. Before benchmarking anything, confirm which version you're getting and which license it carries, because the model page already points to a newer GLM-5.3 checkpoint.

5. MiMo-V2.6 from Xiaomi

Specifications

Xiaomi's current series is MiMo-V2.6, after V2.5 was retired from its platform on 21 October 2026. MiMo-V2.6-Pro has 1.02 trillion parameters with 42 billion active and a one-million-token context, and its model card lists four kinds of input it accepts (text, image, video and audio). V2.6-Flash is smaller. It has 309 billion parameters with 15 billion active.

Best fit

MiMo suits teams building coding agents or tool-using assistants that want an MIT license and a smaller active footprint than DeepSeek V4-Pro or Kimi K3.

Benefits

Xiaomi lists V2.6 under MIT, and for the V2.5 release it stated that commercial use and fine-tuning need no additional authorization. Flash's 15 billion active parameters keep serving costs low.

Caveats

The Pro model card bars military uses and malicious cyber activity and asks for human oversight of high-risk autonomous actions, terms that sit alongside the MIT license and belong in the legal read. Version life is also short. V2.5 entered public testing on 23 April 2026 and was retired about six months later.

6. Doubao-Seed-2.0 from ByteDance

Specifications

ByteDance released Doubao-Seed-2.0 on 14 February 2026 and followed it with Seed 2.1 on 23 June 2026. Both are available both in the Doubao app and through Volcano Engine, ByteDance's cloud API service. The Pro variant is benchmarked against GPT-5.2 and Gemini 3 Pro.

Best fit

It suits cost-sensitive teams that want a hosted model and are comfortable working with ByteDance's cloud.

Benefits

ByteDance says Doubao 2.0's token pricing is about an order of magnitude lower than comparable models. If that figure holds up, the model is a candidate for high-volume work such as classification or summarization, where small per-token savings add up.

Caveats

Doubao has no open weights. Every prompt goes to ByteDance's cloud, so security and legal need to sign off on where the data lives before anyone tests the model with real data from the business. ByteDance also reports leading results on multimodal benchmarks as well as math and coding, and those claims are its own. Check Volcano Engine's current pricing and account rules for your region first.

7. Manus 2.0, the agent built on top of models

Specifications

Manus is a general-purpose AI agent that plans and carries out multi-step jobs, such as researching a market and writing up the report. Butterfly Effect, a company founded in China and based in Singapore, launched it as an invitation-only beta on 6 March 2025, and Manus 2.0 and the Cue app followed on 28 September 2026. It runs on Cascade, Manus's own execution system. In Cue, each agent gets its own email address and phone number, and even a mobile wallet.

Best fit

Manus suits business users who want to hand off a whole research job. The agent does the browsing and writes up what it finds while you get on with something else.

Benefits

Tasks run in the cloud and don't need the user's machine to stay on. Manus also reports that Cascade cost 32% less to run than its previous system in one tested configuration, a number worth treating carefully, because Manus hasn't said how it ran the test and nobody has independently verified it. One configuration is thin evidence. Our first look at the Manus AI agent explains how it plans and executes tasks.

Caveats

Ownership has been unstable. Meta announced a $2 billion acquisition of Manus in December 2025, and on 27 April 2026 China's National Development and Reform Commission blocked it, less than five months later.

CNBC reported in October 2026 that Manus had resumed independent operations. The same report said it raised more than $500 million in a round led by Boyu Capital and IDG Capital. Teams building on Manus should keep track of who owns it.

How DeepSeek R1 changed the cost of AI models

In January 2025, DeepSeek R1 showed that an open reasoning model could compete with a leading US model on a small training budget. DeepSeek states that R1 performs comparably to OpenAI o1 on reasoning tasks, with math and code among them. Its R1 paper in Nature, published on 17 September 2025, put the final training run at $294,000, using 512 Nvidia H800 chips for 80 hours, and that number is easy to misread. It covers the final run only.

The training method mattered as much as the hardware. DeepSeek trained an early version, R1-Zero, with large-scale reinforcement learning and no supervised fine-tuning first (the model was rewarded for reaching correct answers and never shown human-written worked solutions to copy). Our original DeepSeek-R1 breakdown covers the launch-era claims.

Markets reacted within days. On 27 January 2025, according to Reuters, Nvidia shares fell just under 17%, erasing $592.7 billion of market value that day, and the wider Nasdaq index dropped 3.1%. Investors read the cost figures as a threat to ever-larger GPU orders.

What lasted, in our view, is price pressure. Because R1 and V4 are MIT-licensed, any cloud provider can host them and compete on price, and rivals now pitch cost directly, the way ByteDance does with Doubao. For any workload where an open model clears the quality bar, the hosting bill now matters more than the model fee.

What to check before evaluating a Chinese AI model

License terms and data location settle more of these evaluations than benchmark scores do. Do the usual third-party checks, then these five.

• Read the license file itself. MIT (DeepSeek V4, R1, GLM-5.2, MiMo-V2.6) and Apache 2.0 (Qwen3.8-27B) allow commercial use with few conditions, while the Kimi K3 License, the Qwen3.8-Max license and MiMo's use restrictions need a legal read.
• Decide where inference runs. Self-hosted open weights keep prompts inside your own cloud region. A hosted API in another jurisdiction may conflict with GDPR duties or customer contracts, and the only way to know is to confirm where requests are processed and stored.
• Size hardware by total parameters and budget compute by active ones. For Kimi K3, that means memory for all 2.8 trillion weights, while per-token compute tracks only the 104 billion active parameters.
• Test on your own prompts. Most benchmark figures in this guide come from the labs themselves, and a few hundred real requests from your product are a better guide than any leaderboard.
• Track version and vendor stability. Pin model versions in production, and record who owns the vendor, since the Manus acquisition was announced and blocked within five months.

Arbisoft's generative AI team checks models for bias, privacy risks, unintended outputs and performance. For a Chinese model, we'd add the license read and the data-location decision to that same evaluation and settle both before a pilot starts.

Chinese AI models in 2027

These are predictions for 2027, built on what the labs shipped in 2026. It's an outlook, and any of them could turn out wrong.

• We expect version churn to speed up. After a V4 preview in April 2026, DeepSeek shipped a production V4-Pro in August and two Flash updates by October. We expect point releases every few months across most labs, and re-testing will become a standing task for any team running one of these models in production.
• We expect split licensing to spread to more flagship weights. Alibaba already ships Qwen3.8-27B under Apache 2.0 while keeping its 2.4T model under a separate license, and Moonshot ties Kimi K3 to a $20 million revenue threshold. Our guess is that more labs will follow, keeping small models permissive and putting conditions on the largest.
• We expect one-million-token context to stop setting models apart. DeepSeek V4, Kimi K3, GLM-5.2 and MiMo-V2.6 all offer it today, and once every serious model reads a million tokens, no buyer will pick one for that feature alone. Competition will probably move to agent reliability and tool use, where GLM-5.2 and MiMo-V2.6 already aim.
• Ownership review will join model review. The Manus case showed that a Chinese regulator can block a foreign acquisition months after it's announced, and we expect procurement teams to respond by adding ownership and cross-border checks to every AI vendor assessment.

Frequently asked questions

Are Chinese AI models free for commercial use?

Many are. DeepSeek V4 and R1, GLM-5.2 and MiMo-V2.6 ship under MIT, and Qwen3.8-27B uses Apache 2.0. Kimi K3 and the 2.4T Qwen3.8 model use custom licenses with conditions for large companies or products, and Doubao-Seed-2.0 is a paid hosted service with no downloadable weights at all.

Is it safe to send company data to a Chinese AI model?

Self-hosted open weights carry the lowest data risk, because prompts and outputs stay on infrastructure you control. Hosted APIs need more care. Before sending anything sensitive to one, confirm where data is processed and stored and check that against your GDPR obligations and what your customer contracts say about where their data may go.

Which Chinese AI model should a team test first?

Start with DeepSeek V4 or GLM-5.2 (both MIT-licensed) and test it against your current provider. Add Kimi K3 if long documents or images drive the workload.

What happened to Manus AI after the Meta deal?

China blocked it. Meta announced a $2 billion acquisition in December 2025, and China's National Development and Reform Commission prohibited the foreign investment on 27 April 2026. Manus resumed independent operations, raised more than $500 million in a round led by Boyu Capital and IDG Capital, and released Manus 2.0 on 28 September 2026.

Getting help evaluating Chinese AI models

For most teams, the practical first step is DeepSeek V4 or GLM-5.2, both permissively licensed, tested against the model already in production on a sample of real prompts. Bring in Kimi K3 when the work is heavy on lengthy files and images. Hosted Doubao comes last, after data residency is cleared.

Arbisoft's generative AI development team builds and fine-tunes models, then handles the integration work. The team also measures performance and checks each model for bias and privacy risks, with a separate round of testing for unintended outputs. A short proof of concept on your own data will show whether the savings hold before any migration starts.