Best Chinese AI Models 2026: MiniMax, Kimi, GLM-5, Qwen, DeepSeek Compared

Chinese models now hold 61% of OpenRouter token volume. Compare MiniMax M2.5, Kimi K2.5, GLM-5, DeepSeek V3.2, and Qwen3 on benchmarks, pricing, and use cases.

AfricanAI Team 10 min read

As of February 24, 2026, Chinese AI models account for 61% of all token consumption on OpenRouter: the most-used AI API aggregator in the world. The top three most-used models on the platform are all Chinese: MiniMax M2.5 at #1 with 2.45 trillion tokens in a single week, Kimi K2.5 at #2 with 1.21 trillion tokens, and GLM-5 at #3 with 780 billion tokens (Dataconomy).

This is not a gradual trend, it is a structural shift. Chinese labs went from roughly 1% of global AI market share in late 2024 to approximately 15% by November 2025. By February 2026, they hold the majority of developer token usage on the largest neutral platform.

This comparison covers the five strongest Chinese open-weights models available today: MiniMax M2.5, Kimi K2.5, GLM-5, DeepSeek V3.2, and Qwen3-235B. Every benchmark number in this article comes from official lab releases or third-party evaluators, no extrapolation.

Why Chinese AI models now matter

Three things changed simultaneously in the 2025–2026 period:

Pricing pressure. Chinese models have systematically undercut Western proprietary pricing by 5–20x. This is not subsidized dumping, the models are genuinely more efficient, largely due to MoE architectures that activate only a fraction of total parameters per inference step. At scale, the cost difference is decisive.

Benchmark parity. As recently as mid-2024, the best Chinese open models lagged frontier proprietary models by 10–15 percentage points on coding and reasoning benchmarks. That gap has closed. Multiple Chinese open-weights models now score above Claude Sonnet on SWE-Bench and GPQA, with competitive Chatbot Arena ratings from human evaluators.

Open weights. All five models in this comparison release weights publicly, some under MIT (GLM-5), others under modified MIT with attribution. Self-hosting at frontier quality is now possible for teams with GPU access. This removes API latency, vendor lock-in, and per-token cost entirely for teams that can afford inference infrastructure.

The developer community has noticed: programming tasks grew from 11% of OpenRouter token volume in 2025 to over 50% of total volume, and Chinese models now dominate that category (OpenRouter State of AI).

The models

MiniMax M2.5

Released February 12, 2026. A 230B total parameter MoE model with 10B active per token. Holds the #1 OpenRouter position by volume. Context window: 205K tokens. License: modified MIT.

The defining characteristic is the combination of frontier coding performance (80.2% SWE-Bench) at the lowest price in this comparison ($0.30/$1.20 input/output per million tokens). A Lightning variant offers approximately 100 tokens per second at $0.30/$2.40 (MiniMax).

Kimi K2.5

Released January 27, 2026. A 1T total parameter MoE model with ~32B active per token. The only natively multimodal model in this comparison (text, image, video). Context window: 256K tokens. License: modified MIT.

The defining characteristic is Agent Swarm mode: up to 100 parallel sub-agents, 1,500 simultaneous tool calls, 3–4.5x faster execution. This produces the highest HLE with Tools score of any model (50.2%) and a BrowseComp Agent Swarm score of 78.4%, above GPT-5.2 Pro (Kimi K2.5 Blog).

GLM-5

Released February 11, 2026. A 744B total parameter MoE model with 40–44B active per token (256 experts, top-8 activation). Context window: 200K input / 128K output. License: full MIT, no attribution required.

The defining characteristic is the best hallucination calibration of any model: a score of -1 on the AA-Omniscience Index, representing a 56% reduction in hallucinations versus its predecessor and the best result across all models tested, including proprietary systems (Artificial Analysis). It also ranks #1 among open-weights models on the AA Intelligence Index.

DeepSeek V3.2

Released late 2025. A MoE model trained on Huawei Ascend chips with a focus on mathematical reasoning. AIME 2026 score of 96.0%, which surpasses GPT-5 High (94.6%) on the same benchmark (InfoQ). SWE-Bench Verified: approximately 67.8%. GPQA Diamond: approximately 79.9%.

Pricing: $0.28 input / $0.42 output per million tokens, the cheapest in this comparison. Cache hit input: $0.028/M (DeepSeek API docs).

Qwen3-235B

Released by Alibaba's Qwen team. A 235B total parameter MoE model with 22B active (Apache 2.0 license, the most permissive in this comparison). GPQA Diamond: 81.1%. AIME 2025: approximately 92.3% in thinking mode. LiveCodeBench v5: 70.7%.

Available via API at approximately $0.20–$1.20 per million input tokens depending on context length, and $1.00–$6.00 per million output tokens. The Apache 2.0 license means it can be used in any commercial product without attribution requirements.

Benchmark comparison

The table below shows performance on the four most widely used third-party benchmarks. All figures are from official lab announcements or Artificial Analysis evaluations. Empty cells indicate the benchmark was not publicly reported by the lab or evaluator at time of writing.

Model SWE-Bench Verified GPQA Diamond AIME 2026 HumanEval
MiniMax M2.5 80.2% 84.8% , ,
Kimi K2.5 76.8% 87.6% , 99.0%
GLM-5 77.8% 86.0% 92.7% ,
DeepSeek V3.2 ~67.8% ~79.9% 96.0% ,
Qwen3-235B , 81.1% 92.3%* ,

*Qwen3 AIME figure is for the 2025 exam in thinking mode; GLM-5 and DeepSeek V3.2 AIME figures are for the 2026 exam.

Coding (SWE-Bench): MiniMax M2.5 leads at 80.2%, followed by GLM-5 (77.8%) and Kimi K2.5 (76.8%). DeepSeek V3.2 is lower at ~67.8% but is still competitive for many coding tasks.

Science reasoning (GPQA Diamond): Kimi K2.5 leads at 87.6%, with GLM-5 (86.0%) and MiniMax M2.5 (84.8%) close behind. Qwen3-235B (81.1%) and DeepSeek V3.2 (~79.9%) are somewhat lower.

Mathematics (AIME): DeepSeek V3.2 leads at 96.0% on AIME 2026, surpassing GPT-5. GLM-5 scores 92.7% on the same exam.

Agentic/research (HLE with Tools): Only Kimi K2.5 has published this number, 50.2% with Agent Swarm, which is the current world record on Humanity's Last Exam with tools enabled.

Hallucination: GLM-5 is in a category of its own with a -1 score on AA-Omniscience. No other model in this table (or beyond) has matched it.

Pricing comparison

All prices are per 1 million tokens as of February 2026.

Model Input Output Cache Hit Input License
DeepSeek V3.2 $0.28 $0.42 $0.028 Open (MIT-like)
MiniMax M2.5 $0.30 $1.20 , Modified MIT
MiniMax M2.5 Lightning $0.30 $2.40 , Modified MIT
Kimi K2.5 $0.60 $3.00 $0.10 Modified MIT
Qwen3-235B $0.20–$1.20* $1.00–$6.00* , Apache 2.0
GLM-5 $1.00 $3.20 $0.20 MIT (full)
Claude Sonnet 4.6 (ref) $3.00 $15.00 , Proprietary

*Qwen3-235B uses tiered pricing by context length range.

Cheapest output: DeepSeek V3.2 at $0.42/M is the most economical for output-heavy workloads, roughly 35x cheaper than Claude Sonnet on output. However, its SWE-Bench score is lower.

Best output quality per dollar: MiniMax M2.5 at $1.20/M output with an 80.2% SWE-Bench score offers the best coding performance per dollar of output cost.

Most permissive license: Qwen3-235B (Apache 2.0) and GLM-5 (MIT) allow unrestricted commercial use without attribution. MiniMax M2.5 and Kimi K2.5 require attribution under their modified MIT licenses.

Use case guide

Best for high-volume coding / software engineering MiniMax M2.5. It has the highest SWE-Bench score (80.2%) and the lowest effective cost for output-heavy coding tasks. The Lightning variant at ~100 t/s is the right choice for real-time code completion pipelines.

Best for autonomous research agents and complex multi-step pipelines Kimi K2.5. The Agent Swarm capability (100 parallel sub-agents, 1,500 tool calls) is architecturally unique among open models. The 50.2% HLE with Tools score and 78.4% BrowseComp result confirm that swarm-based research outperforms single-model approaches on hard tasks.

Best for factual accuracy and knowledge-intensive applications GLM-5. The -1 Omniscience Index score is the defining data point. For legal research, medical information, scientific literature synthesis, or any application where confident wrong answers are unacceptable, GLM-5's calibration advantage is decisive. The full MIT license is a bonus for enterprise deployments.

Best for mathematics and quantitative reasoning DeepSeek V3.2. A 96.0% AIME 2026 score, above GPT-5 High, combined with the cheapest pricing in this comparison ($0.28/$0.42) makes it the best choice for math tutoring, quantitative finance tools, engineering calculation support, and statistical analysis.

Best for teams needing unrestricted open-source use Qwen3-235B (Apache 2.0) or GLM-5 (MIT). If your product cannot accommodate an attribution requirement, these are the only two models in this comparison with fully unrestricted commercial licenses. GLM-5 has better benchmark scores; Qwen3-235B has better licensing terms and a more established deployment ecosystem through Alibaba's infrastructure.

Best for multimodal applications (image, video, text) Kimi K2.5. It is the only natively multimodal model in this comparison. MiniMax M2.5, GLM-5, DeepSeek V3.2, and Qwen3-235B are primarily text-only (some have separate vision model variants, but the unified architecture in Kimi K2.5 handles mixed-modality inputs more coherently).

Best for creative writing and content generation Kimi K2.5. At #16 globally on the lechmazur creative writing benchmark (score: 8.068), it is the strongest Chinese open-weights model for prose quality. Its predecessor K2-0905 held position #7 (8.331), the best creative writing score ever recorded for a Chinese-developed model. For content platforms, marketing tools, or any product where output prose quality is evaluated by end users, K2.5 is the clear choice among models in this comparison.

Open-source considerations

All five models release weights publicly, but the licenses differ in ways that matter for production deployment:

Apache 2.0 (Qwen3-235B): Completely unrestricted. Use commercially, modify, redistribute, and sublicense without attribution. The most permissive option for enterprise products.

MIT (GLM-5): Also completely unrestricted for commercial use. No attribution required. Simple and clear for legal review.

Modified MIT (MiniMax M2.5, Kimi K2.5): Commercial use is permitted, but attribution is required when the weights are incorporated into products. The specific attribution requirement should be reviewed in each model's license file before production deployment.

DeepSeek V3.2: Uses a DeepSeek-specific open license that permits commercial use. Consult the license text directly for specific restrictions.

Self-hosting on your own infrastructure eliminates API vendor risk and removes per-token cost entirely, relevant for teams running inference at high volume. For a 1T-parameter model like Kimi K2.5, full-precision inference requires substantial GPU memory (roughly 2TB for BF16 weights), but quantized variants run on smaller clusters.

For teams without GPU infrastructure, all five models are available through multiple API providers including the labs' own endpoints, NVIDIA NIM, and third-party aggregators.

Verdict

The choice among these five models depends almost entirely on your primary use case:

  • Coding at scale, cost-sensitive: MiniMax M2.5, best SWE-Bench, lowest cost for quality.
  • Agentic research, multi-step pipelines: Kimi K2.5, no open model matches its swarm architecture.
  • Factual reliability, knowledge tasks: GLM-5, best hallucination calibration of any model, full MIT license.
  • Mathematics, quantitative work: DeepSeek V3.2, AIME 2026 record, cheapest pricing.
  • Unrestricted commercial license, general use: Qwen3-235B, Apache 2.0, strong general performance.

The 61% OpenRouter market share tells you something important: developers have already run the cost-benefit analysis and made their choice. At 5–20x lower prices than proprietary models with equivalent or better benchmark performance, the question for most teams is no longer "should I use a Chinese open model?", it is "which one?"

All five models are available on AfricanAI: run them side by side in your browser without API key management.

Sources: