The best Chinese open-source LLMs in 2026

The best Chinese open-source LLMs right now are MiMo-V2.6-Pro, Qwen3.8-27B and GLM-5.3, out of 17 we have reviewed. Each is ranked by our rating, with its price, our verdict and who it suits.

By Andrew Miller · Updated Oct 26, 2026 · How we review

How we ranked: by our rating out of 5, from reviews built on each product's own pricing, documentation and policies plus independent benchmarks where they exist, each checked on the date shown in the review. We take no affiliate commission.

Also ranked: the best open-source LLMs

1. 1
   M

   MiMo-V2.6-Pro

   Licence and data: Weights under MIT. Xiaomi's MiMo privacy policy says data is stored in Singapore and deleted or anonymised after its retention period; the Desktop app trains on inputs unless you opt out. Third-party gateways report that Xiaomi's API does not train on traffic, which Xiaomi's API terms should confirm.

   Running it: $0.435/$0.87 per million tokens with a 99% cache discount and half-price batch. 1,000 agent tasks of 500K input and 30K output tokens cost about $244 uncached; Flash would be about $78.

   Price: Open weights under MIT. Xiaomi API: Pro $0.435 input / $0.87 output per 1M tokens, cached input $0.004; Flash $0.14 / $0.28; Pro UltraSpeed $4.35 / $8.70; Batch API half price.

   Read our MiMo-V2.6-Pro review
   4.5/5
2. 2
   Q

   Qwen3.8-27B

   Licence and data: Apache 2.0: commercial use with no revenue caps or acceptable-use annex. Qwen Cloud's listing publishes no retention or training terms, and its sample endpoint is in Singapore (ap-southeast-1).

   Running it: Self-hosted: 4-bit 16.5 GB, 8-bit 29 GB, FP8 31 GB, BF16 about 56 GB of weights before KV cache. Hosted: Qwen Cloud $0.50/$3.00 per 1M tokens; OpenRouter providers range from about $0.04/$1.25 to $0.45/$3.20.

   Price: Free weights under Apache 2.0. Qwen Cloud: $0.50 input / $3.00 output per 1M tokens ($0.10 cached). Third-party hosts on OpenRouter from about $0.04 / $1.25.

   Read our Qwen3.8-27B review
   4.5/5
3. 3
   G

   GLM-5.3

   Licence and data: Z.ai's API is run from Singapore and its data processing terms say it processes customer data on the customer's behalf and does not store API inputs or outputs, which are processed in real time; an explicit statement on training with API prompts is not published. Self-hosting keeps data in-house.

   Running it: Z.ai lists $1.40/$4.40 per 1M tokens with cached input at $0.26 and cache storage free for a limited time. Artificial Analysis puts it at $2.01 per index task, against $0.72 for GPT-6.1 Sol. Self-hosting means about 753 GB of FP8 weights plus cache memory.

   Price: Open weights under a modified MIT licence. Z.ai API: $1.40 input / $0.26 cached / $4.40 output per 1M tokens; GLM-5.3-Flash $0.15 / $0.50.

   Read our GLM-5.3 review
   4.4/5
4. 4
   D

   DeepSeek-V4-Pro

   Licence and data: Open weights are MIT-licensed with no commercial or revenue restrictions. DeepSeek's privacy policy (February 2026) says data is stored in China, lists model training among its uses of user input, and offers opt-out only by emailing privacy@deepseek.com.

   Running it: On the API, 1B input tokens a month at 70% cache hits plus 100M output tokens costs about $820 at peak rates or $410 off-peak. Self-hosting needs roughly 865 GB for weights alone, which by our arithmetic means at least an eight-GPU H200 server (1.1 TB of HBM) and more for full 1M-token contexts.

   Price: Free MIT-licensed weights. DeepSeek API: $0.66 input / $1.98 output per 1M tokens off-peak, $1.32 / $3.96 at peak; cache hits from $0.022.

   Read our DeepSeek-V4-Pro review
   4.3/5
5. 5
   G

   GLM-5.3-Flash

   Licence and data: Z.ai's API data terms say prompts and outputs are processed in real time and not stored, and customer data is generally processed in Singapore. The terms do not mention training either way, and no SOC 2 or ISO certification is published. Z.ai serves the model on Chinese-made accelerators.

   Running it: On Z.ai, 2B input tokens (80% cached) and 200M output tokens a month cost about $208; GLM-5.3 would be about $1,856. Self-hosting means roughly 330 GB of FP8 weights plus KV cache, which is more than four 80 GB GPUs can hold, so plan on an eight-GPU node.

   Price: Open weights under MIT. Z.ai API: $0.15 input / $0.50 output per 1M tokens ($0.03 cached); FlashX $0.37 / $1.25.

   Read our GLM-5.3-Flash review
   4.3/5
6. 6
   K

   Kimi K3

   Licence and data: Moonshot's privacy policy allows training on user content with no API exemption described, keeps data while the account is active and stores it in Singapore; no SOC 2 or ISO certification is named. The licence exempts internal use from its revenue and display clauses.

   Running it: $3/$15 per million tokens, $0.30 cached, with extra charges for 5-minute or 1-hour cache writes. 1,000 agent tasks of 500K input and 30K output tokens cost about $1,950 uncached; Fireworks lists one endpoint at $2.10/$10.50.

   Price: Open weights under a modified MIT licence. Kimi API: $3 input / $15 output per 1M tokens; cached input $0.30.

   Read our Kimi K3 review
   4.3/5
7. 7
   D

   DeepSeek-V4.1-Flash

   Licence and data: Weights under MIT with no use restrictions. On DeepSeek's service, the privacy policy (updated 10 February 2026) says data is stored in China, kept while the account exists and may train models, with an opt-out on request; no separate API retention terms are published.

   Running it: 10 billion input and 1 billion output tokens a month costs about $2,100 off-peak or $4,200 at peak on DeepSeek; cached input is $0.003 per million off-peak. Search and Fetch add $0.75 per 1,000 successful calls each, so an agent making 20 searches per task across 10,000 tasks pays $150 a month for search; free through harnesses until 31 October 2026, 16:00 UTC. Self-hosting needs over 765 GB of GPU memory for the weights alone, and DeepInfer's deployment notes put DeepSeek's smallest efficient unit at 64 prefill plus 64 decode GPUs.

   Price: Free MIT-licensed weights. DeepSeek API: $0.15 input / $0.60 output per 1M tokens off-peak, $0.30 / $1.20 at weekday peak; cache hits $0.003 to $0.006.

   Read our DeepSeek-V4.1-Flash review
   4.2/5
8. 8
   H

   Hy4 preview

   Licence and data: Weights under Apache 2.0 with no extra conditions. Tencent Cloud's China-site terms say it will not use inputs to improve its own models; retention is not stated publicly. Models from other labs on TokenHub follow their own terms.

   Running it: $0.834/$2.501 per million tokens on TokenHub. 1,000 agent tasks of 500K input and 30K output tokens cost about $492 uncached. Self-hosting FP8 needs eight B300-class GPUs.

   Price: Open weights under Apache 2.0. Tencent Cloud TokenHub: ¥6 input / ¥18 output per 1M tokens, or $0.834 / $2.501 internationally, cached input $0.042. OpenRouter from about $0.75 / $2.25. Hy3: $0.13 / $0.53.

   Read our Hy4 preview review
   4.2/5
9. 9
   Q

   Qwen3.8-Flash-Next

   Licence and data: Qwen Community License 1.0: commercial use allowed; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service and AI coding or office assistant businesses at any size. Qwen Cloud's data-security docs cover encryption and in-region storage but publish no retention period or training policy.

   Running it: Hosted: $0.15 input and $0.47 output per million tokens and $0.016 for cache hits; Batch is offered with no published discount. Self-hosted: roughly 360 GB of BF16 or 180 GB of FP8 weights, so a multi-GPU H100 or H200 server before KV cache.

   Price: Open weights under the Qwen Community License 1.0 (separate licence for hosting and AI-assistant businesses). Qwen3.8-Flash API: $0.15 input / $0.47 output per 1M tokens; cache hits $0.016.

   Read our Qwen3.8-Flash-Next review
   4.2/5
10. 10
    Q

    Qwen3.8-Max

    Licence and data: Licence allows commercial use; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service or AI work assistant businesses above $50 million yearly revenue; internal use is exempt. Qwen Cloud's data-security docs cover encryption and in-region storage but publish no retention period or training policy.

    Running it: 10 million input and 2 million output tokens a day, a busy internal assistant, costs about $32 a day or roughly $960 a month at list price; explicit cache reads at $0.17 cut the cost of repeated context by over 90%.

    Price: Open weights under the Qwen3.8-Max licence. Qwen Cloud API: $2 input / $6 output per 1M tokens; cache reads $0.17 (explicit) or $0.25 (implicit).

    Read our Qwen3.8-Max review
    4.2/5
11. 11
    S

    Step 5 Preview

    Licence and data: Weights under the StepFun Model License 1.0: display rule above 100M users or $20M monthly revenue, commercial licence for device makers and MaaS businesses over $20M a year, internal use exempt. StepFun's API terms allow training on Startup and early-access programme traffic; terms for other paid traffic are not explicit.

    Running it: $1.00/$2.70 per million tokens with 90% off cached input. 1,000 agent tasks of 500K input and 30K output tokens cost about $581 uncached. Self-hosted FP8 needs four B300 or eight H200 GPUs; BF16 needs eight B300 or sixteen H200.

    Price: Open weights under the StepFun Model License 1.0. StepFun API: $1.00 input / $2.70 output per 1M tokens, 90% off cached input; same price on OpenRouter and DeepInfra.

    Read our Step 5 Preview review
    4.2/5
12. 12
    L

    Ling-3.1-flash

    Licence and data: Weights under MIT, with a public training-content summary. inclusionAI's API privacy and retention terms were not found on its model or OpenRouter pages; self-host for sensitive data.

    Running it: At $0.30/$0.90, 1,000 agent tasks of 500K input and 30K output tokens cost about $177 uncached, before its above-average reasoning output. At OpenRouter's listed price, about a quarter of that.

    Price: Open weights under MIT. inclusionAI API: $0.30 input / $0.90 output per 1M tokens, cached $0.06, per Artificial Analysis; OpenRouter lists $0.075 / $0.225.

    Read our Ling-3.1-flash review
    4.0/5
13. 13
    M

    MiniCPM5-2B

    Licence and data: Apache 2.0 for code and weights, with no revenue cap or acceptable-use addendum. Self-hosted only, so prompts stay on your hardware; no Hugging Face inference provider hosts it.

    Running it: No hosted price exists, so cost is hardware: about 5 GB of BF16 weights, 2.68 GB at Q8_0 and 1.56 GB at Q4_K_M, which a single consumer GPU or a laptop holds with room for context. OpenBMB publishes no throughput figures to size a fleet.

    Price: Free weights and training data under Apache 2.0.

    Read our MiniCPM5-2B review
    4.0/5
14. 14
    M

    MiniMax-M3.1-Flash

    Licence and data: Licence: commercial use needs a "Built with MiniMax M3.1-Flash" credit and a one-time notice to MiniMax; products over $20 million in yearly revenue need prior written authorisation; military use is prohibited. MiniMax's API retention and training terms were not published on the pages we could read.

    Running it: 10 billion input and 1 billion output tokens a month is about $2,100 on MiniMax's API, with cached input at $0.015 per million; reasoning at max effort inflates the output side. Self-hosting needs about 250 GB of GPU memory for the MXFP8 weights before KV cache, within an eight-GPU H100 80 GB node.

    Price: Open weights under the MiniMax community licence. MiniMax API per 1M tokens, flat at any prompt length: $0.15 input, $0.60 output, $0.015 cached input.

    Read our MiniMax-M3.1-Flash review
    4.0/5
15. 15
    M

    MiniMax-M3

    Licence and data: Licence: commercial use needs a "Built with MiniMax M3" credit and a notice to MiniMax; products over $20 million in annual revenue need prior written permission; no military use. MiniMax's API terms keep ownership with you but let it use inputs and generated content to develop and improve its services; no opt-out or retention period is published.

    Running it: 1M support replies a month at 2,000 input and 300 output tokens each is about $960 on the standard tier, less with cached system prompts. Self-hosting at BF16 needs roughly 850 GB of GPU memory before KV cache: an eight-GPU H200 or B200 node, or more than one H100 node.

    Price: Open weights under the MiniMax community licence. API per 1M tokens on prompts up to 512K: $0.30 input, $1.20 output, $0.06 cached; double above 512K; priority tier 1.5x.

    Read our MiniMax-M3 review
    3.9/5
16. 16
    Q

    Qwen3.5-397B-A17B

    Licence and data: Apache 2.0 allows commercial use, fine-tuning and redistribution with no revenue or user limits. Self-hosting keeps prompts in-house; on OpenRouter it is not available through EU in-region routing, and each host sets its own retention.

    Running it: OpenRouter hosts charge $0.39 to $0.75 input and $2.34 to $4.50 output per 1M tokens; hosted Qwen3.5-Plus is $0.30/$1.80. Self-hosting needs about 406 GB for FP8 (an eight-GPU 80 GB server), 214 GB at 4-bit or 180 GB at 3-bit.

    Price: Free weights under Apache 2.0; Qwen3.5-Plus on Alibaba Cloud and third-party hosting billed separately.

    Read our Qwen3.5-397B-A17B review
    3.6/5
17. 17
    X

    Xing4.0-29B-A4B

    Licence and data: Apache 2.0: commercial use and fine-tuning allowed with no revenue cap or acceptable-use addendum. Self-hosted, nothing leaves your servers; the OpenRouter endpoint is run by a third party, Phala, under its own terms.

    Running it: BF16 weights are about 62GB (31.2B parameters including embeddings), one 80GB H100; FP8 halves that; a community GGUF build at 11.6GB fits a 16GB card. The only hosted option, OpenRouter's free endpoint, is capped at 20 requests a minute and 50 a day (1,000 a day once an account has bought $10 of credits), so production traffic means self-hosting.

    Price: Free weights under Apache 2.0. No first-party API; OpenRouter lists a free endpoint (hosted by Phala) under its free-model request caps.

    Read our Xing4.0-29B-A4B review
    3.6/5