DeepSeek and Chinese AI models explained: Qwen, Kimi, GLM, MiMo, and MiniMax

Parameter count is a poor guide to how a Chinese AI model will behave in production. The largest open-weight model available today, Moonshot’s 2.8-trillion-parameter Kimi K3, completes only 13.3% of Toloka Arena’s agentic tasks reliably across five runs. Xiaomi’s MiMo V2.6 Pro, at about a third of the size, scores 58.7% at a fraction of the cost per task, and a 27-billion-parameter Qwen model you can run on a single workstation outscores both Kimi K3 and DeepSeek’s current API model.

That gap matters because Chinese labs now set the pace for open-weight AI. DeepSeek, Alibaba’s Qwen team, Moonshot, Z.ai, Xiaomi, and MiniMax all shipped new flagship models between July and October 2026, almost all with downloadable weights, 1 million token context windows, and API prices well below those of US frontier models. For teams that want to self-host, fine-tune, or simply cut inference costs, these families are now the default shortlist rather than an afterthought.

Unlike the Claude, GPT, and Gemini families, this is not one vendor’s tier ladder. It is six independent labs with different licences, release rhythms, and strengths, plus a lifecycle problem of their own: model names are frequently re-pointed to newer checkpoints without a change on the caller’s side.

This guide explains what each family offers as of October 2026, compares their models on size, context, licence, and price, shows how they score on private agentic benchmarks, and covers the deployment and compliance questions that come with choosing a Chinese model.

What are Chinese AI models?

Chinese AI models are large language models developed by labs based in mainland China. The families most relevant to global developers in 2026 are DeepSeek, Qwen from Alibaba, Kimi from Moonshot AI, GLM from Zhipu AI (which operates internationally as Z.ai), MiMo from Xiaomi, and MiniMax. Each publishes a hosted API and, in most cases, the model weights themselves.

Three traits set the group apart from the leading US families. First, open weights are the norm rather than the exception, so teams can download a checkpoint and run it on their own infrastructure. Second, nearly all current flagships are sparse mixture-of-experts models, which activate only a small share of their parameters for each token and keep inference costs low relative to their total size. Third, API prices are aggressive: several capable models cost under $1 per million output tokens.

“Open weights” is not the same as open source, and the difference is practical. DeepSeek and Xiaomi publish under the MIT licence and Qwen’s smaller models under Apache 2.0, while Moonshot, Z.ai, and MiniMax use custom licences that can add attribution requirements or commercial conditions for large deployments. Our guide to open-source vs closed-source LLMs covers why that distinction matters.

How Chinese AI models evolved

The turning point was January 2025, when DeepSeek released R1, an open reasoning model under the MIT licence that matched leading proprietary models on several reasoning benchmarks at a much lower reported training cost. Our analysis of the data behind DeepSeek’s success looks at how it was built. Alibaba followed with the Qwen 3 family in April 2025, Moonshot with the trillion-parameter Kimi K2 in July 2025, and Z.ai, MiniMax, and Xiaomi with their own open MoE models over the following year.

2026 brought scale and speed. DeepSeek previewed its V4 family in April, with a 1 million token context window as the default across its services. Moonshot launched Kimi K3 in July at 2.8 trillion parameters, Alibaba released the 2.4-trillion-parameter Qwen 3.8 Max in August and opened its weights, the first time a Max-class Qwen model became downloadable, and Z.ai shipped GLM-5.3 the same month. Xiaomi followed with MiMo V2.6 in September.

The pace has not slowed. In October alone, DeepSeek updated its main API model (V4.1 Flash 1017), MiniMax released M3.1 Flash with open weights four days after launch, and Moonshot shipped Kimi K3.1 Pro with weights promised by early November. Release notes are now measured in weeks, and that cadence shapes how these models should be evaluated and deployed.

Chinese AI model families: DeepSeek, Qwen, Kimi, GLM, MiMo, and MiniMax

All six families now converge on similar headline specifications: trillion-scale sparse models, 1 million token context, and agentic coding as the primary target. The differences that matter are in how each lab packages, prices, and licenses its models. Prices below are first-party list prices per million input and output tokens unless noted, and they change often.

DeepSeek: V4.1 Flash and V4 Pro

DeepSeek currently serves two models through its API. DeepSeek V4.1 Flash, released on September 10, 2026 and updated to the 1017 checkpoint on October 17, is the default. It is a 552-billion-parameter MoE model with native image input, a 1 million token context window, up to 384,000 output tokens, and MIT-licensed weights. DeepSeek says the 1017 update raised its Terminal-Bench 4.0 score from 31.2 to 42.3 while shortening outputs and task times, at an unchanged price.

The larger DeepSeek V4 Pro, generally available since August 13, is a 1.6-trillion-parameter model with 49 billion active parameters and three reasoning effort levels. DeepSeek has referred to a V4.1 Pro but has not dated it.

Two operational details are unique to DeepSeek. It bills by time of day: since August 2026, off-peak rates are half the peak rate, so V4.1 Flash costs $0.15 and $0.60 off-peak and $0.30 and $1.20 at peak, and V4 Pro $0.66 and $1.98 off-peak. And it reroutes old model names to new checkpoints: deepseek-v4-flash now silently serves V4.1 Flash, so the model answering your requests can change without any change in your code. Record the model version alongside every evaluation result.

Qwen: Qwen 3.8 Max and Qwen 3.8 27B

Alibaba’s Qwen family has the broadest size range of any Chinese lab, from small dense models for edge devices to its largest flagship. Qwen 3.8 Max, released on August 3, 2026, is a 2.4-trillion-parameter MoE model with 95 billion active parameters, text, image, and video input, close to 1 million tokens of context, and up to 131,000 output tokens. Its API price is $2 per million input tokens and $6 per million output tokens, and an upgraded 0902 snapshot followed in September with coding and agent improvements.

The bigger shift is openness. Alibaba had previously kept its Max-class flagships proprietary; in August it made the Qwen 3.8 Max weights downloadable alongside Qwen 3.8 27B, a natively multimodal dense model under Apache 2.0 with a 262,000 token context window that is built to run on a single workstation or high-end laptop. The 27B model is the realistic self-hosting option; at 2.4 trillion parameters, the Max checkpoint is a multi-node datacenter artifact. Check the model card for the licence on each checkpoint before deploying.

The 27B size class is also where fine-tuning pays off most visibly. We measured what one epoch of our OTS enterprise RL data does for Qwen 3.5 27B, the previous generation of the same model line.

Kimi: Kimi K3 and Kimi K3.1 Pro

Moonshot AI’s Kimi K3, launched on July 16, 2026 with weights released on July 27, was the largest open-weight model released to date: a 2.8-trillion-parameter MoE with native vision and a 1 million token context window. Its weights ship natively in 4-bit MXFP4 precision, which makes the checkpoint smaller than its parameter count suggests but still far beyond consumer hardware.

On October 23, Moonshot released Kimi K3.1 Pro, a heavily post-trained K3, at $2 per million input tokens and $10 per million output tokens, a third below K3. A Highspeed variant runs two to three times faster at double the price. Weights are promised by November 4, and Moonshot says it removed cyber-related training data from the open release. Independent testing by Artificial Analysis notes that K3.1 Pro is very verbose, so its cost per task is closer to that of pricier models than its token price suggests.

Kimi K3 is released under the Kimi K3 licence, a modified MIT licence that requires a separate agreement for large model-as-a-service resellers. Our Arena write-up on why Kimi K3 solves the task once, not five times in a row explains the gap between its launch benchmarks and its reliability scores.

GLM: GLM-5.3 and GLM-5.3 Flash

Z.ai released GLM-5.3 on August 14, 2026, a post-training upgrade on the GLM-5 base aimed at long-horizon coding and agents, with a 1 million token context window and text-only input. It is available through OpenRouter at around $1.40 per million input tokens and $4.40 per million output tokens. GLM-5.3 Flash, a smaller natively multimodal model with 320 billion total and 18 billion active parameters, followed on August 26 under the MIT licence.

GLM-5.3 is notable for how its weights were released. Z.ai held them for roughly two weeks of extra safety review after the model showed unusually strong vulnerability-discovery ability, then published them at the end of August under a bespoke GLM-5.3 licence rather than the MIT licence used for earlier GLM-5 models.

MiMo: MiMo V2.6 Pro and V2.6 Flash

Xiaomi is the newest of the six to compete at the frontier. MiMo V2.6 Pro, released on September 21, 2026 under the MIT licence, is a 1.02-trillion-parameter MoE model with 42 billion active parameters, a 1 million token context window, and text, image, video, and audio input. At $0.435 per million input tokens and $0.87 per million output tokens, it is one of the cheapest capable models available.

MiMo V2.6 Flash costs $0.14 and $0.28, and a Pro UltraSpeed variant trades a tenfold price increase for much faster output. Xiaomi retires older checkpoints quickly: the V2.5 models left the API on October 21, about a month after V2.6 launched.

MiniMax: MiniMax M3 and M3.1 Flash

MiniMax’s M-series has shipped open weights since M1. MiniMax M3, released on June 1, 2026, is a 428-billion-parameter MoE with 23 billion active parameters and a 1 million token context window. MiniMax M3.1 Flash launched on the API on October 19 at $0.15 per million input tokens and $0.60 per million output tokens, with weights published four days later. It accepts text, image, and video input and targets coding, agents, and computer use.

The weights come under the MiniMax Community License, which requires attribution in commercial deployments and prior written authorisation from MiniMax once a deployer’s annual revenue passes $20 million.

Other Chinese model families

ByteDance (Doubao and Seed), Tencent (Hunyuan), Baidu (ERNIE), and StepFun also build capable models. They matter most inside their own ecosystems and the Chinese domestic market, and appear less often in global developer deployments than the six families above, so they are outside the scope of this comparison.

DeepSeek vs Qwen vs Kimi vs GLM vs MiMo vs MiniMax

The current flagships side by side show how similar the headline specifications have become, and how much licences and prices still differ.

Sources: vendor changelogs, model cards, and pricing pages, checked October 26, 2026. GLM-5.3 price is via OpenRouter; first-party Z.ai pricing may differ.

How Chinese models score on Toloka Arena

Specifications describe what a model is; reliability describes what it does. Toloka Arena evaluates models on private, multi-turn tool-use tasks with strict business rules and reports a composite pass^5 score, the probability that all five independent runs of a task succeed, alongside the average inference cost per task. Because the tasks are never published, no model can have trained on them, which matters when most launch benchmarks are public and can leak into training data.

Source: Toloka Arena composite leaderboard, updated October 22, 2026. Kimi K3.1 Pro and MiniMax M3.1 Flash were released after this update and are not yet ranked.

Three things stand out. MiMo V2.6 Pro is the strongest Chinese model on the board, sits on the cost-reliability frontier, and is the cheapest model of all 31 to clear 50% pass^5, landing within five points of Gemini 3.8 Flash at about 40% of the cost per task. Qwen 3.8 27B, a model small enough to self-host on one machine, outscores DeepSeek V4.1 Flash, Kimi K3, and GLM-5.3, though it has slipped off the frontier since newer entrants such as Reflection Beam now deliver more reliability for less. And Kimi K3 and GLM-5.3, both of which led coding and agent benchmarks at launch, complete only about one task in eight reliably across five runs.

The lesson is the same one the US families teach, only sharper: launch benchmarks measure whether a model can solve a task, and pass^5 measures whether it solves it every time. Neither parameter count nor price predicts the second. Evaluate shortlisted models on your own representative tasks, and rerun the evaluation whenever a vendor re-points a model name to a new checkpoint.

Chinese AI models vs Claude, GPT, and Gemini

The comparison with US families is less about raw capability than about trade-offs. On the October 22 Arena board, the best Chinese model ranks 14th of 31, behind 13 models from US labs, but ahead of budget tiers such as Claude Haiku 5.5 (28.4%), GPT-6 Luna (39.1%), and Grok 4.7 (37.5%), as well as Mistral Large 4 (39.2%), and at a much lower cost per task than the models above it. Our LLM leaderboard overview tracks the wider picture.

Price and cost per task

Token prices for Chinese models are typically a fraction of US frontier prices: MiMo V2.6 Pro’s output tokens cost less than a twentieth of Claude Opus 5.5’s. But token price is not task price. Verbose models generate far more output tokens per task, and reasoning-heavy configurations multiply that again. Independent testing put Kimi K3.1 Pro’s cost per task at roughly the same level as its predecessor despite a one-third price cut, because it writes more. Cost per completed task, measured on your own workload, is the number to compare.

Openness and licences

Open weights are the clearest structural advantage. Teams can self-host for data control, fine-tune on proprietary data, quantise for cheaper hardware, or distil a large model into a smaller one, none of which is possible with Claude, GPT, or Gemini. Google’s Gemma, OpenAI’s gpt-oss line, and Meta’s open models compete here, but Chinese labs currently release their largest and most capable checkpoints openly, which US frontier labs do not. Licence terms still need a legal read: MIT and Apache 2.0 are permissive, while the Kimi, GLM-5.3, and MiniMax licences add conditions that matter at scale.

Deployment, compliance, and data residency

Where the model runs determines most of the compliance question. Calling a Chinese lab’s first-party API sends prompts and data to that provider, which many organisations cannot accept for regulated or sensitive data, and the DeepSeek app has drawn regulatory scrutiny in several countries. Self-hosting the open weights, or using a third-party host in your own region, keeps data under your control, which is why open weights and enterprise adoption are closely linked.

Two further considerations belong in any procurement review. The White House science office has accused Moonshot of training Kimi K3 partly by distilling outputs from US models, which Moonshot denies, and teams should track how that question develops. And cyber capability is now a release gate on both sides: Z.ai delayed GLM-5.3’s weights for safety review and Moonshot stripped cyber data from Kimi K3.1’s open release, mirroring the restricted-access programmes US labs use for their most capable cyber models. Independent red teaming before deployment is prudent whichever family you choose.

What comes next for Chinese AI models?

The near-term roadmap is visible. Moonshot has promised Kimi K3.1 Pro’s weights by November 4, DeepSeek has referred to a V4.1 Pro without a date and describes V4.1 Flash as the smallest model in a new architecture family, and MiniMax’s M3.1 line has so far shipped only its Flash model.

The broader trends are price compression, faster cadence, and a shift toward agent harnesses. DeepSeek now optimises its models for specific coding agents, including its own DeepSeek Harness, and MiniMax and Moonshot ship their own coding tools. Expect more models tuned for particular agent frameworks, more frequent checkpoint updates behind stable names, and continued pressure on US API prices from the open-weight end of the market.

How human data improves open-weight models

Open weights change what teams can do with human data. With a closed API, the vendor decides what the model learns; with open weights, a team can fine-tune on its own domain, its own tools, and its own definition of a correct answer. That is often where a mid-ranked base model becomes the best model for a specific workflow.

Reinforcement learning on realistic enterprise tasks is particularly effective for agentic reliability. We showed how a 4B agent model can move from 6% to 71% task success on enterprise tool use in under a day, and the same approach applies to the 27B and larger checkpoints covered here. The constraint is data: tasks that reflect real business rules, verifiable rewards, and expert judgement on the cases automated checks miss.

Toloka provides that data. Teams can request our off-the-shelf RL and evaluation datasets, commission custom LLM training and fine-tuning data and coding data, or run expert model evaluation and AI safety testing before deployment.

Frequently asked questions

What are the main Chinese AI models in 2026?

The six families most relevant to global developers are DeepSeek (V4.1 Flash and V4 Pro), Alibaba’s Qwen (Qwen 3.8 Max and Qwen 3.8 27B), Moonshot’s Kimi (Kimi K3 and K3.1 Pro), Z.ai’s GLM (GLM-5.3), Xiaomi’s MiMo (MiMo V2.6 Pro), and MiniMax (M3 and M3.1 Flash). All offer hosted APIs, and most publish open weights.

What is the latest DeepSeek model?

The latest is DeepSeek V4.1 Flash 1017, an update released on October 17, 2026 to the V4.1 Flash model from September. It is served through the deepseek-flash API name, with MIT-licensed weights, a 1 million token context window, and native image input. The larger DeepSeek V4 Pro remains available, and DeepSeek has mentioned a V4.1 Pro without giving a release date.

Are Chinese AI models open source?

Most are open-weight rather than strictly open source. DeepSeek and MiMo use the MIT licence and Qwen 3.8 27B uses Apache 2.0, which are permissive. Kimi, GLM-5.3, and MiniMax use custom licences that can require attribution, a separate agreement for large resellers, or authorisation above a revenue threshold. Read the licence on each checkpoint before commercial use.

DeepSeek vs Qwen: which should I use?

It depends on cost and deployment. DeepSeek V4.1 Flash is among the cheapest capable APIs, especially at off-peak rates. Qwen offers the wider size range, including a 27B model that self-hosts on a single machine and scored higher than DeepSeek V4.1 Flash on Toloka Arena’s agentic tasks in October. For agentic workloads, test both on your own tasks, and check whether the current Arena entry reflects DeepSeek’s October 17 update.

Which Chinese AI model is best for agentic tool use?

On Toloka Arena’s October 22, 2026 leaderboard, MiMo V2.6 Pro is the most reliable Chinese model at 58.7% pass^5 and about $0.74 per task, ranking 14th of 31 models, followed by Qwen 3.8 Max 0902 at 50.3% and Qwen 3.8 27B at 41.0%. Kimi K3 and GLM-5.3 score about 13% despite strong launch benchmarks. Newer releases such as Kimi K3.1 Pro and MiniMax M3.1 Flash were not yet ranked, so check the current standings before choosing.

Is it safe to use Chinese AI models with business data?

The main risk depends on where the model runs. Sending data to a Chinese lab’s first-party API may conflict with data-residency or sector rules, while self-hosting open weights or using a regional third-party host keeps data under your control. Either way, review the licence, test the model for unsafe or unexpected behaviour on your own tasks, and keep a record of the exact checkpoint you approved.

Why do Chinese models score lower on Arena than on their launch benchmarks?

Launch benchmarks are usually public and single-run, so they measure whether a model can solve a task at least once, and public test sets can leak into training data. Toloka Arena uses private tasks and requires success in all five independent runs. Models that are capable but inconsistent, or that break business rules in long multi-turn interactions, score much lower under that standard.

From model choice to system design

Chinese open-weight models have turned model selection into an architecture decision. A single system can route high-volume, low-risk work to a cheap model such as MiMo V2.6 Flash or DeepSeek V4.1 Flash, reserve a stronger proprietary model for the hardest steps, and fine-tune a self-hosted smaller model for the workflow that matters most.

Those decisions are only as good as the evaluation and training data behind them. Teams need representative tasks, verifiable rewards, and human judgement on real failure modes, ideally from datasets built for agentic training and evaluation. Otherwise, the model that looks best on a launch chart may not be the one that performs best in production.

Related reading

Toloka Arena: independent evaluation of agentic intelligence

Knowledge distillation

How frontier labs build pre-training datasets

R1 is not on par with o1, and the difference is qualitative, not quantitative