Chinese AI Models Compared: DeepSeek, Qwen, Kimi, Doubao, and Ernie

DeepSeek dominates Western coverage of Chinese AI, but it is only one part of a wider field. Qwen, Kimi, ByteDance’s Seed models behind Doubao, ERNIE and Zhipu’s GLM family differ on openness, licensing, access and workload fit, and the newest documented releases, including DeepSeek V4.1-Flash, Qwen3.8, Kimi K3, Seed2.0 and GLM-5.3, have shifted the comparison even since mid-2026.

The result is a routing map rather than a single ranking. Independent measurements, vendor-reported numbers and list prices are labeled as such throughout, because in this ecosystem the provenance of a number changes how much weight it should carry.

---

Why the Chinese AI Ecosystem Deserves a Map

The DeepSeek-R1 paper, published January 2025, reported MATH-500 at 97.3, slightly ahead of OpenAI o1-1217’s 96.4, and AIME 2024 at 79.8 against o1’s 79.2. Those results changed the price-performance conversation, but they did not describe the whole Chinese ecosystem.

By spring 2026, Chinese models accounted for 41% of Hugging Face downloads (State of Open Source report) and had overtaken U.S. models in monthly and overall downloads (see Groundy’s report breakdown). The same report (Hugging Face) records different corporate strategies: Alibaba’s Qwen family has more than 113,000 derivative models, Baidu went from zero Hub releases in 2024 to more than 100 in 2025, and ByteDance increased its release count by eight to nine times.

Those are different kinds of strength. Qwen’s advantage is ecosystem breadth. DeepSeek combines open weights with aggressive API pricing. Kimi emphasizes long context and agent swarms. GLM pairs a 1M-token context with MIT-licensed GLM-5.2 weights. ByteDance’s Seed2.0 is a production API line on Volcano Engine, and Baidu’s ERNIE 4.5 is a ten-variant Apache 2.0 family.

---

The Competitors at a Glance

---

DeepSeek: Open Weights, Cheap Off-Peak API Access

DeepSeek-V3 established the architectural template: 671B total parameters with 37B active per token, Mixture-of-Experts routing and Multi-head Latent Attention to compress key-value representations (DeepSeek-V3 Technical Report). The R1 training story needs more precision than it usually receives. DeepSeek-R1-Zero was the model trained through large-scale reinforcement learning without supervised fine-tuning as a preliminary step. DeepSeek-R1 itself added thousands of cold-start examples and used a multi-stage pipeline with two supervised fine-tuning stages and two reinforcement-learning stages, according to the R1 paper.

The V4 line is larger and more expensive than the V3 generation. Artificial Analysis reports V4 Pro at 1.6T total and 49B active parameters, with V4 Flash at 284B total and 13B active (Artificial Analysis). V4 Pro scored 52 on its Intelligence Index, second among open-weights models behind Kimi K2.6’s 54 in that evaluation, and led the open-weights field on GDPval-AA agentic work with 1554 against K2.6’s 1484. The same evaluation measured hallucination rates of 94% for V4 Pro and 96% for V4 Flash: when the models lacked an answer, they nearly always supplied one anyway.

V4.1-Flash, published September 10, 2026, is a different design point from the retired V4 Flash. It is a multimodal Mixture-of-Experts model with a 552B backbone and a Causal Encoder-Decoder layout (V4.1-Flash model card): 8B parameters activate per token during prefill and 16B during decode, cutting the global KV cache footprint to 890 bytes per token, roughly a quarter of V4-Flash’s. DeepSeek’s own comparison table reports V4.1-Flash at 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, ahead of the larger V4 Pro on both; those are vendor-run numbers at maximum reasoning effort, not independent measurements. The repository and weights are MIT-licensed.

DeepSeek’s API documentation, as of September 21, 2026, lists two models: deepseek-flash, backed by DeepSeek-V4.1-Flash, and deepseek-v4-pro. Retired deepseek-v4-flash requests are now served by V4.1-Flash and billed at Flash rates. Both current models carry a 1M context length with a 384K maximum output, and only the Flash tier accepts image input. According to the current price grid, V4 Pro costs $0.66 per million input tokens off-peak and $1.32 at peak, with output at $1.98 and $3.96, while V4.1-Flash costs $0.15 and $0.30 for input and $0.60 and $1.20 for output. Cache-hit Flash input falls to $0.003 off-peak and $0.006 at peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays; weekends are off-peak.

Use case fit: mathematics, code and reasoning workloads where MIT-licensed V4 weights or very low off-peak API costs matter. The hallucination measurements make V4 a poor default for unsupervised factual answering, and batch workloads should be scheduled outside weekday peak windows where possible.

---

Qwen: Breadth First, License Second

Qwen’s clearest verified advantage is adoption breadth. The Qwen family has more than 113,000 derivative models on Hugging Face (State of Open Source report), rising above 200,000 when models tagged Qwen are included. That derivative ecosystem gives teams more fine-tunes and deployment options than a single flagship score would suggest.

On capability, Qwen3.6 Max Preview tied DeepSeek V4 Pro at 52 on the Artificial Analysis Intelligence Index in the reporting summarized by DeepLearning.AI. Alibaba then unveiled Qwen3.7-Max on May 20, 2026 as its flagship agent model, designed for coding, debugging, office automation and long-horizon tasks. In an internal test summarized by TechNode, Alibaba reports more than 1,000 tool calls and up to 35 hours of sustained autonomous execution; those are vendor-reported results, not independent measurements.

The newest generation is Qwen3.8. The open Qwen3.8-27B (model card) is a dense 27B vision-language model with a hybrid layout mixing Gated DeltaNet linear attention and gated full attention, 262,144 tokens of native context extensible to 1M via YaRN, and vendor-reported scores of 61.7 on SWE-bench Pro under the Claude Code harness and 73.0 on Terminal-Bench 2.1 under the Terminus harness. A hosted version with 1M context by default is announced as coming soon. On the hosted side, Alibaba Cloud’s Model Studio catalog lists qwen3.8-max among its text models and also serves third-party Chinese models, including DeepSeek V4 Pro, DeepSeek V4.1-Flash, GLM-5.2, GLM-5.3, Kimi K3 and MiniMax-M2.5, for its customers (Alibaba Cloud Model Studio). For the architectural tradeoffs inside the Qwen3.6 generation, see Groundy’s dense-versus-MoE analysis.

Licensing should not be inferred from the Qwen family name. Terms can change between releases and checkpoints, so the model card or repository LICENSE file is the controlling source for any specific checkpoint.

Use case fit: teams that value a large derivative ecosystem, multiple deployment paths and Alibaba’s agent-model roadmap.

---

Kimi: Long Context, Agent Swarms, and a 2.8T Flagship

Kimi K2.6 is a 1T-parameter mixture-of-experts vision-language model with 32B active parameters, a 256K-token input window and up to 98K output tokens. It supports tool use, web search, native INT4 quantization, a “preserve thinking” mode and an agent-swarm mode that can coordinate up to 300 subagents over 4,000 steps (DeepLearning.AI).

In the May 2026 reporting, K2.6 scored 54 on the Artificial Analysis Intelligence Index, ahead of DeepSeek V4 Pro and Qwen3.6 Max Preview at 52, but behind the leading proprietary models. Its hallucination rate, as measured by Artificial Analysis, came in at 39.26% (DeepLearning.AI), a substantial improvement over K2.5’s 64.6% and near Claude Opus 4.7’s 36.18%. Reported K2.6 API pricing is $0.95 per million input tokens, $0.16 per million cached input tokens and $4.00 per million output tokens.

Moonshot’s newer release is Kimi K3, a 2.8T-parameter mixture-of-experts model with 104B active parameters built on Kimi Delta Attention and Attention Residuals (Kimi K3 repository). It activates 16 of 896 experts per token, adds native vision through a MoonViT-V2 encoder, and ships a 1M-token context window with quantization-aware training in MXFP4; Moonshot calls it the world’s first open 3T-class model. Moonshot’s own comparison table, which mixes its runs with scores cited from Artificial Analysis and public leaderboards as of July 23, 2026, reports K3 at 88.3 on Terminal-Bench 2.1, 42.0 on SWE-Marathon (the best score in that row; Claude Opus 4.8 follows at 40.0) and 1,686 Elo on GDPval-AA v2, behind Claude Fable 5’s 1,747 (Kimi K3 repository). Treat that table as vendor documentation. API access is documented at platform.kimi.ai with OpenAI- and Anthropic-compatible endpoints.

K3’s weights and code ship under the Kimi K3 License. It grants broad usage rights but adds two commercial conditions (Kimi K3 License): operators of a “model as a service” business whose aggregate revenue exceeds $20 million over any consecutive 12 months must sign a separate agreement with Moonshot, and products above 100 million monthly active users or $20 million in monthly revenue must display “Kimi K3” prominently. Internal use and use through Moonshot’s own products or certified partners are exempt. K2.6’s terms, as reported by DeepLearning.AI, carry a similar attribution threshold, and each Kimi release ships its own license text.

Use case fit: long-document analysis, multi-step tool use and agentic coding. K2.6 remains the highest-scoring open-weights model on the cited May 2026 Intelligence Index snapshot and has published per-token pricing; K3 is the newer flagship with a 1M window, and the K3 scores quoted above come from Moonshot’s comparison table.

---

Doubao and Seed2.0: ByteDance’s Production Line on Volcano Engine

ByteDance’s release activity rose sharply in 2025, by eight to nine times on Hugging Face according to the Spring 2026 ecosystem report. The primary documentation of the current line is ByteDance’s official Seed2.0 launch announcement and the Seed2.0 repository from its Seed team.

The launch announcement is explicit about the Doubao relationship: Seed’s LLM series has supported consumer products with hundreds of millions of users, such as Doubao, and the Seed2.0 Pro and Code models have launched in the Doubao app and TRAE, with the API for the full series available on Volcano Engine. Doubao is the product surface; Seed2.0 is the model line behind it.

Seed2.0 is a production-oriented series with three general-purpose agent tiers, Pro, Lite and Mini, plus a dedicated Code model, built, the team says, on a 500x growth in daily token usage (Seed2.0 repository). Its capability claims are the vendor’s own: leading multimodal results on MMSIBench, MotionBench and VideoMME, gold-medal results in ICPC, IMO and CMO tests with FrontierSci scores exceeding GPT-5.2 in some scenarios, and token pricing roughly one order of magnitude below comparable top-tier industry models (official launch). The announcement also cites public preference standings, 6th on the LMSYS Chatbot Arena text leaderboard overall and 3rd on the vision leaderboard as of February 16, 2026. The repository is a cookbook for the Volcano Engine API rather than a weights release: its Apache-2.0 license covers the example code, so code licensing and model licensing are separate questions, and model terms and per-token pricing should be requested from ByteDance directly.

Use case fit: teams already on Volcano Engine, or teams evaluating ByteDance’s API who can obtain official pricing and terms for their region. The vendor’s cost claim, roughly a tenth of comparable flagship token costs, is the number to test in a pilot.

---

ERNIE 4.5: Apache 2.0 Weights With a PaddlePaddle Toolchain

Baidu’s open-release posture changed quickly: Hugging Face (State of Open Source report) records zero Hub releases in 2024 and more than 100 in 2025, and names Baidu among the organizations that shifted decisively toward open releases.

The anchor release is ERNIE 4.5, a ten-variant family documented in the official ERNIE repository: mixture-of-experts language models at 300B total with 47B active and 21B with 3B active, vision-language models up to 424B total that accept text, image and video input, and a 0.3B dense model, with 128K context on the language tiers. All ERNIE 4.5 models, weights included, are provided under Apache License 2.0 with commercial use permitted, distributed on Hugging Face and AI Studio, and trained on PaddlePaddle at a reported 47% Model FLOPs Utilization for the largest language model (ERNIE repository).

Baidu’s results are vendor measurements (ERNIE repository): ERNIE-4.5-300B-A47B-Base is reported to surpass DeepSeek-V3-671B-A37B-Base on 22 of 28 benchmarks; ERNIE-4.5-21B-A3B-Base is reported to outperform Qwen3-30B-A3B-Base on BBH and CMATH at roughly 70% of the total parameters; and the VL-424B model’s thinking mode is claimed to narrow or surpass the gap to OpenAI o1 on MathVista, MMMU and VisualPuzzle. Thinking variants arrived later in 2025: ERNIE-4.5-21B-A3B-Thinking in September and ERNIE-4.5-VL-28B-A3B-Thinking with ERNIEKit v1.5 in November. The ERNIEKit and FastDeploy toolkits cover SFT, DPO, quantization-aware training and OpenAI-compatible serving.

Use case fit: teams that want Apache 2.0 weights with an industrial fine-tuning and deployment toolchain, particularly for Chinese-language or multimodal workloads. The Apache 2.0 terms are stated for the ERNIE 4.5 models.

---

GLM-5.2 and GLM-5.3: MIT Weights Meet a Post-Training Step Change

Zhipu announced GLM-5.2 on June 13, 2026, with availability across GLM Coding Plan tiers and open-source availability following that week (release coverage). The GLM-5 repository now presents GLM-5.3 as the newer flagship. GLM-5.3 uses the same base model as GLM-5.2, with its claimed coding and long-horizon gains coming from post-training.

GLM-5.2 is a 744B-A40B model with a 1M-token context window. Its IndexShare sparse-attention design reuses one indexer across every four sparse-attention layers, reducing per-token FLOPs by 2.9x at 1M context, while an improved MTP layer supports speculative decoding. The GLM-5.2 model card states that the weights carry an MIT open-source license with “no regional limits,” and now points readers to GLM-5.3-BF16 as the newer version.

Zhipu’s vendor-reported GLM-5.2 results include SWE-bench Pro at 62.1, Terminal-Bench 2.1 at 81.0 under the Terminus-2 harness, AIME 2026 at 99.2, HMMT November 2025 at 94.4, GPQA-Diamond at 91.2 and HLE at 40.5 on the text-only subset (GLM-5.2 model card). The same card’s comparison cells for Qwen3.7-Max, DeepSeek V4 Pro and Claude Opus 4.8 are Zhipu’s measurements of other labs’ models, making this a vendor-reported comparison.

For GLM-5.3, Zhipu claims a 50% improvement over GLM-5.2 on its internal Z.ai Code Bench and open-weights state of the art on Terminal Bench 3.0 and Agents’ Last Exam (GLM-5 repository). Zhipu also reports that cyber capability grew faster than expected during post-training, with GLM-5.3 topping CyberGym for vulnerability discovery and more than doubling GLM-5.2 on exploitation benchmarks. GLM-5.3-Flash is separately trained at 320B-A18B, combining sparse and linear attention with Manifold-Constrained Hyper-Connections and a 30T-token multimodal pre-training corpus (GLM-5 repository).

Hosted access is subscription-based through the GLM Coding Plan. Z.ai’s developer documentation advertises plans starting at $18 per month for Lite, with weekly credit allowances of 10,000 for Lite, 60,000 for Pro and 140,000 for Max, and off-peak usage charged at half the standard credit rate. All plans now support GLM-5.3 and GLM-5.3-Flash, while GLM-5.2 and GLM-5.1 requests are automatically routed to GLM-5.3. The MIT-licensed GLM-5.2 weights remain the self-hosting option, with no commercial revenue or attribution thresholds.

See the GLM-5.2 launch analysis and the licensing comparison for more detail.

---

Benchmark Comparison: What the Evidence Actually Shows

The independent cross-lab anchors are limited but useful. Artificial Analysis and DeepLearning.AI put Kimi K2.6 at 54 on the Intelligence Index, with DeepSeek V4 Pro and Qwen3.6 Max Preview tied at 52 (Artificial Analysis; DeepLearning.AI). DeepSeek V4 Pro leads the cited GDPval-AA comparison at 1554 against Kimi K2.6’s 1484. DeepSeek-R1’s own paper reports MATH-500 at 97.3 and AIME 2024 at 79.8 (R1 paper).

The following cells all come from Zhipu’s GLM-5.2 model card, including the measurements of competitors:

* The card’s footnote marks Claude Opus 4.8’s HLE 49.8 as a full-set figure, while the GLM-5.2 score is the text-only subset the card reports by default.

These numbers do not make GLM-5.2 the independently verified leader. They show where Zhipu positions it on Zhipu’s selected harnesses. DeepSeek’s V4.1-Flash card adds a second vendor-run cross-lab table, reporting Terminal-Bench 2.1 at 90.6 for V4.1-Flash against 88.3 for Kimi K3 and 88.2 for GLM-5.3 at maximum reasoning effort (V4.1-Flash model card). Moonshot’s K3 table and Baidu’s ERNIE 4.5 claims are likewise vendor-published. These tables use different tests, harnesses and publication dates, so they should be read as separate evaluations.

---

API Pricing at a Glance

Prices as of September 2026.

---

Choosing the Right Model

The selection decision still turns on three questions: do you need weights, can you use the hosted service from your jurisdiction, and which workload dominates?

• Self-host a large reasoning model: DeepSeek V4 Pro under MIT. Weigh the 94% hallucination rate measured by Artificial Analysis before using it for factual answers without retrieval or review.
• Self-host long-context coding weights: GLM-5.2. Its MIT terms and 1M context are clear, while SWE-bench Pro 62.1 and Terminal-Bench 81.0 are vendor-reported anchors from the model card.
• Use a broad derivative ecosystem: Qwen. The 113,000-plus derivative count is verified, and the Qwen3.8 generation adds a dense open 27B option with vendor-reported coding scores close to the frontier.
• Run long-context API work: Kimi K2.6 at 256K input with published pricing, or K3 with a 1M window through Moonshot’s documented API.
• Minimize a straightforward API bill: DeepSeek V4.1-Flash off-peak at $0.15 input and $0.60 output per million tokens, according to the current pricing grid.
• Build on ByteDance or Baidu: Seed2.0 is the documented option for Volcano Engine customers, with vendor cost claims worth piloting; ERNIE 4.5 offers Apache 2.0 weights and a full PaddlePaddle toolchain for self-hosters.
• Compare against a Western price reference: Claude Fable 5 costs $10/$50 per million tokens, far above DeepSeek or Kimi list prices, according to Anthropic. It is a capability and price reference, not a substitute for any open-weight model here.

For coding-specific routing, pair this comparison with Groundy’s AI code generation benchmarks and the best AI models for OpenClaw. The practical answer is rarely “the highest-ranked model.” It is the model whose benchmark evidence, license and access path all survive procurement review.