Chinese AI Models Q2 2026: 10-Provider Landscape Report
Q2 2026 report on 10 Chinese AI providers, from Xiaomi and Qwen to DeepSeek and Kimi: OpenRouter token share, list prices, licenses and enterprise fit.
We rechecked this report against archived OpenRouter data and vendor documentation. The Chinese token share is now stated as the 59.6% its own table sums to, rather than “more than 45%”; the year-earlier baseline (12–16%, not under 2%), GLM-5’s active parameters (40B), several prices, the DeepSeek IMO/IOI attribution and MiniMax’s licensing are corrected, and unsupported claims about training hardware are removed.
For later gateway-usage evidence, see our September 26 report on Chinese models’ share of OpenRouter tokens. That article distinguishes OpenRouter and Vercel traffic from the overall AI market and identifies the dates and limits of each observation. The provider landscape and figures below remain the original Q2 snapshot; they are not a current market-share measurement.
Providers covered
Chinese labs’ OpenRouter token share, early April 2026
Xiaomi vs OpenAI OpenRouter token share
Xiaomi OpenRouter token share, early April 2026
Key Takeaways
Chinese AI labs held nearly 60% of OpenRouter tokens in the week to April 3, 2026, up from 12–16% for DeepSeek and Qwen combined in April 2025. Xiaomi alone had nearly three times OpenAI’s token share. This is the Q2 2026 landscape report: who the ten leading providers are, what they ship, how they price it, and how to evaluate them for production workloads.
The shift is not a benchmark story. Chinese models do not yet lead the Artificial Analysis Intelligence Index — on the version shown on OpenRouter’s leaderboard on April 3, MiMo-V2-Pro ranked #10 despite being the #1 model by usage. The shift is a cost, availability, and developer-choice story. Free-preview access, 1M token context windows, and per-token list prices from about 2.5x to more than 20x below US frontier models have pulled a large share of OpenRouter’s coding and tool-calling traffic toward Chinese models.
Data source: OpenRouter Rankings, read on April 3, 2026. Provider shares come from OpenRouter’s token share by model author for the week beginning March 29, which was still in progress; the Internet Archive’s snapshots from April 2 and April 3 show that week at 17.2T and 21.4T tokens and bracket the token totals below. Model totals, the category and tool-call charts and the Intelligence Index ranks come from a later April 3 snapshot; the model leaderboard covers a trailing week, so the two windows differ. OpenRouter counts prompt plus completion tokens routed through its gateway only, using each provider’s own tokenizer, so a token from one provider is not directly comparable with a token from another. These are token shares on one platform, not market shares, and they measure usage, not model quality. For the full ranking breakdown see our April 2026 OpenRouter rankings analysis.
Q2 2026 Landscape at a Glance
Ten providers account for almost all Chinese-model tokens on OpenRouter in early Q2 2026: Xiaomi, Alibaba, Z.ai (Zhipu), DeepSeek, Moonshot AI, MiniMax, StepFun, ByteDance, Baidu, and Tencent. Baichuan, Yi, Xunfei, and KwaiKAT operate as second-tier niche players.
| Rank | Provider | Weekly tokens | Share | Flagship model |
|---|---|---|---|---|
| 1 | Xiaomi | 4.21T | 21.1% | MiMo-V2-Pro |
| 2 | Alibaba (Qwen) | 2.77T | 13.9% | Qwen 3.6 Plus |
| 3 | MiniMax | 1.62T | 8.1% | MiniMax M2.7 |
| 4 | Z.ai (Zhipu) | 1.12T | 5.6% | GLM-5 / GLM-5 Turbo |
| 5 | DeepSeek | 1.11T | 5.6% | DeepSeek V3.2 |
| 6 | StepFun | 1.07T | 5.3% | Step 3.5 Flash |
| 7 | Moonshot AI | Outside top 10 | Not shown | Kimi K2.5 |
| 8 | ByteDance | Outside top 10 | Not shown | Seed 2.0 (Doubao) |
| 9 | Baidu | Outside top 10 | Not shown | ERNIE 5.0 |
| 10 | Tencent | Outside top 10 | Not shown | HY 2.0 (Hunyuan) |
Source: OpenRouter’s token share by model author, week beginning March 29, 2026, as displayed on April 3. The week was still in progress: archived snapshots from April 2 and April 3 bracket every token total, and each share is within two points of both readings. OpenRouter’s chart names only its ten largest authors; Moonshot, ByteDance, Baidu and Tencent fall in its “others” group and rank in that order on its trailing-week model leaderboard, where Kimi K2.5 had 0.51T tokens. Baidu, Tencent, and Moonshot concentrate usage on domestic Chinese surfaces and partner ecosystems rather than OpenRouter, so their global ranking understates their home-market presence. OpenRouter notes that token counts come from each provider’s own tokenizer, so a token on one row is not directly comparable with a token on another; treat the shares as a usage signal for Western buyers, not a like-for-like measure.
Update, October 2026: most flagships in this table have been succeeded since this snapshot, among them DeepSeek V4, GLM-5.1, Kimi K2.6, MiMo-V2.5-Pro and MiniMax M3. The figures here remain the early-April snapshot.
Mapping models to your stack? Model selection is rarely a single-benchmark decision. Explore our AI Digital Transformation service to translate this landscape into a production-ready architecture.
Xiaomi: The Phone Company Dominating AI Volume
Xiaomi is the story of this report. A consumer electronics company best known for smartphones and smart-home hardware holds 21.1% of OpenRouter weekly tokens, nearly three times OpenAI’s 7.5%. Xiaomi's AI lab shipped three frontier checkpoints between December 2025 and March 2026 under the MiMo brand, and each variant carved out a clear slot in the pricing curve.
Over 1T parameters (42B active), 1M context, $1 input / $3 output per million tokens up to 256K ($2 / $6 beyond). The #1 model on OpenRouter at 4.79T weekly tokens and 25.5% of programming-category tokens.
262K context, $0.40 input / $2 output per million tokens. Unified image, video, and audio architecture in a single checkpoint.
309B parameters (15B active), MIT-licensed weights, 256K context, $0.09 input / $0.29 output per million tokens on OpenRouter.
For deeper technical coverage, see our MiMo-V2-Pro trillion-parameter release guide and the MiMo-V2-Omni omnimodal release guide.
Why Xiaomi won volume
Three decisions drove the rankings. First, Xiaomi built demand before launch: its anonymous test build, Hunter Alpha, topped OpenRouter’s daily chart for several days and passed 1T tokens, and at launch Xiaomi gave developers a week of free API access through OpenClaw, OpenCode, KiloCode, Blackbox and Cline. MiMo-V2-Pro itself was a paid listing on OpenRouter. Second, the 1M context window matched Qwen 3.6 Plus and the 1M windows of GPT-5.4 and Claude Opus 4.6, at a fraction of their price, which suits whole-repo refactors. Third, the pricing gap versus Claude Opus 4.6 is roughly 5x at input and 8x at output, which matters when agent frameworks expand token consumption by an order of magnitude.
Alibaba Qwen: The Scale Leader
Alibaba's Qwen family is the second-ranked provider at 13.9% of OpenRouter tokens and 2.77T weekly tokens, but it is the broadest product line in Chinese AI. Between late January and April 2, 2026, Alibaba shipped at least six named Qwen releases covering coding, reasoning, ultra-cheap inference, omnimodal, and flagship tiers, alongside the open-weight Qwen3.5-397B-A17B in February. Prices below are OpenRouter list prices in early April.
- Qwen 3.6 Plus (Apr 2, 2026) — Flagship for agentic coding. 1M context by default, 65,536-token maximum output on OpenRouter, reasoning and native function calling. Free during the OpenRouter preview; a paid listing at $0.325 / $1.95 had replaced it by April 10. #2 on OpenRouter with 1.64T weekly tokens.
- Qwen 3.5-Omni (late March 2026) — Native omnimodal. Qwen’s launch post and technical report give 256K context, speech recognition in 113 languages and dialects and a Thinker-Talker architecture. Offered through Alibaba Cloud’s API, with no open weights.
- Qwen 3.5 Flash (Feb 24, 2026) — Ultra-cheap high context. 1M tokens, $0.065 input / $0.26 output per million. Top pick for cost-sensitive batch workloads.
- Qwen 3 Coder Next (early Feb 2026) — Coding-specific. 80B parameters (3B active), 256K context, $0.12 input / $0.75 output per million. Built for coding agents and IDE integrations.
- Qwen 3 Max Thinking (late Jan 2026; API snapshot dated January 23) — Reasoning variant. 262K context, $0.78 input / $3.90 output per million. Alibaba compared it with GPT-5.2 Thinking, Claude Opus 4.5 and Gemini 3 Pro on hard problems.
- Qwen 3.5 small series (Mar 2, 2026) — On-device. 0.8B to 9B parameters, Apache 2.0. Qwen’s model card puts the 9B variant at 81.7 on GPQA Diamond, above OpenAI’s open-weight gpt-oss-120b at 80.1.
For current flagship detail, see our Qwen 3.6 Plus 1M-context release guide.
Alibaba also sells Qwen to enterprises through Alibaba Cloud International. For Western teams evaluating Chinese models under compliance constraints, Qwen’s Apache 2.0 open weights and Alibaba Cloud’s published SOC 2 report make it a practical starting point.
Zhipu GLM: Enterprise Enablement
Z.ai — the international brand for Zhipu AI — holds 5.6% of OpenRouter tokens and is positioned toward Chinese enterprise deployment. GLM-5 launched February 11, 2026 with 744B total parameters, 40B active per token, a Mixture-of-Experts architecture with 256 experts, 28.5T pre-training tokens and an MIT license. Zhipu’s technical report describes full-stack inference support for seven Chinese chip platforms, Huawei Ascend among them, which makes GLM-5 a natural fit for Chinese state-owned and export-control-sensitive buyers. The report does not say what hardware trained the model.
- 744B total parameters, 40B active, 200K context window, 128K maximum output.
- 77.8% SWE-bench Verified in Z.ai’s own runs, against 80.9% for Claude Opus 4.5 in the same table.
- $1.00 input / $3.20 output per million tokens on Z.ai’s API, $1.20 / $4.00 for the faster GLM-5 Turbo variant.
- MIT licensed weights — among the most permissive in the Chinese frontier tier.
- GLM-5V-Turbo (April 1, 2026) adds video, image and file input for vision-based coding and agent tasks; Z.ai cites a smaller parameter size but publishes no parameter count.
For the full architectural and benchmark breakdown, see our Zhipu GLM-5 744B MoE release analysis.
Zhipu's positioning is the clearest example of the bifurcation in Chinese AI. While Xiaomi and Alibaba optimize for global developer volume, Zhipu optimizes for Chinese enterprise deployment, domestic-hardware independence, and permissive licensing. For Western self-hosters and researchers, GLM-5’s MIT license sits with the Apache 2.0 Qwen 3.5 open-weight models among the most permissive options; Kimi K2.5 uses a modified MIT license.
DeepSeek: Open-Weight Economics
DeepSeek holds 5.6% of OpenRouter tokens through DeepSeek V3.2, a 685B parameter MoE model released December 1, 2025 with DeepSeek Sparse Attention and DeepSeek’s first integration of thinking into tool use. DeepSeek describes V3.2 as GPT-5-level; its high-compute sibling, V3.2-Speciale, is the model it credits with gold-level results at IMO and IOI 2025 and with surpassing GPT-5. Speciale was served API-only and without tool calls, and these are vendor-run results.
For a detailed walkthrough of the V3.2 release, see our DeepSeek V3.2 and Speciale complete guide.
DeepSeek V4 status: As of April 2026, V4 is expected but not released. Public reporting, citing The Information, points to Huawei Ascend as a primary hardware target. Treat any V4 timeline as unconfirmed until DeepSeek publishes an official release. Update, October 2026: DeepSeek released V4 as a preview on April 24, 2026: V4-Pro (1.6T total, 49B active) and V4-Flash (284B, 13B active), both with 1M context and MIT-licensed weights; V4-Pro reached general availability on August 13. DeepSeek’s release notes do not name the hardware; Huawei said at launch that it had worked with DeepSeek on V4 compatibility across its Ascend line, Reuters reported, but how much of the training used Ascend chips has not been disclosed.
DeepSeek's strategic value is open-weight economics. V3.2 is available for self-hosting, fine-tuning, and quantization under the MIT license, terms that many Western teams find acceptable. When the production decision comes down to "run the model inside my own cloud region or not," DeepSeek is usually on the shortlist alongside GLM-5 and Kimi K2.5.
Moonshot Kimi: Long-Context Agents
Moonshot AI's Kimi K2.5 launched January 27, 2026 with 1 trillion total parameters, 32B active per request, a 256K (262,144-token) context window, open weights under a modified MIT license, and an OpenRouter list price of $0.38 input / $1.72 output per million tokens. The standout feature is Agent Swarm: K2.5 can direct up to 100 sub-agents in parallel across up to 1,500 tool calls, which Moonshot says cuts execution time by up to 4.5x. Moonshot’s own results put the agent swarm ahead of Claude Opus 4.5 on BrowseComp and WideSearch; those are vendor-run evaluations.
Kimi K2.5 is also notable as the base model of Cursor Composer 2, which Cursor’s technical report says scored 73.7% on SWE-bench Multilingual in its own harness at the March 19 launch. That makes K2.5 one of the few Chinese models actively embedded in a Western developer tool's production stack, rather than served only as a standalone API.
For the full agent swarm architecture breakdown, see our Kimi K2.5 agent swarm open-source guide.
MiniMax: Self-Evolving Agentic Workflows
MiniMax holds 8.1% of OpenRouter tokens and is the provider most associated with agentic self-evolution. Its M-series shipped four numbered releases in under six months: M2, M2.1, M2.5 (February 12, 2026, 80.2% SWE-Bench Verified in MiniMax’s runs), and M2.7 (March 18, 2026, 56.22% SWE-Pro, about 10B active of 229B total parameters). At $0.30 / $1.20 per million tokens, M2.7’s list price is about 17x below Claude Opus 4.6 on input and 21x on output. MiniMax M2.7 was #4 on OpenRouter’s April 3 leaderboard at 1.34T weekly tokens with +24% week-over-week growth. Its weights went up on Hugging Face just before this report, under a license that bars commercial use without MiniMax’s written permission.
For deeper coverage on the M2.7 release, see our MiniMax M2.7 agentic coding release guide.
MiniMax's positioning is the inverse of Xiaomi's. Rather than competing on raw context, MiniMax leans into what it calls self-evolution: during M2.7’s development the model updated its own memory, built skills for reinforcement-learning experiments and helped refine its own training process, and an internal version optimized a programming scaffold over more than 100 rounds for a 30% gain. That describes how the model was built, not a model that rewrites itself during a task. For agentic workloads, M2.7 is one of the lowest-priced paid models in OpenRouter’s weekly top ten.
StepFun and Second-Tier Providers
Step 3.5 Flash from StepFun (listed on OpenRouter January 29, 2026, with open weights on Hugging Face from February 1) is the surprise performer outside the top four. A 196B MoE with 11B active per token, 256K context, Apache 2.0 weights, and $0.10 input / $0.30 output per million tokens in the paid tier, its free OpenRouter variant ranked #3 on OpenRouter’s overall weekly model leaderboard at 1.38T tokens. StepFun quotes 100–300 tokens per second in typical use, peaking at 350 for single-stream coding; its model card prices decoding on NVIDIA Hopper GPUs but does not name the training hardware.
Below StepFun, a handful of second-tier providers serve specific niches:
- ByteDance Seed 2.0 (Doubao) — Doubao remained mainland China’s top consumer AI app in December 2025, according to an industry report cited by the South China Morning Post. ByteDance benchmarks the Seed 2.0 Pro variant, released in February, against GPT-5.2 and Gemini 3 Pro in its own evaluations, at a fraction of GPT-5.2’s price. Seed 2.0 Lite and Mini extend the family into cheaper tiers ($0.10-$0.25 per million input on OpenRouter).
- Baidu ERNIE 5.0 — 2.4T-parameter multimodal flagship, launched January 22, 2026 according to the South China Morning Post; Baidu has not said which chips trained it. The same report put monthly active users of Baidu’s AI assistant at 200 million.
- KwaiKAT KAT-Coder-Pro V2 — Coding-specific Kuaishou model, listed on OpenRouter on March 27, 2026. 256K context, $0.30 / $1.20 per million, aimed at enterprise software engineering.
- Baichuan and Yi — Baichuan’s recent open releases are medical models (Baichuan-M3-235B, January 2026); 01.AI has published no new open Yi model since August 2024. Neither has a meaningful OpenRouter presence.
- Xunfei (iFlytek) — Spark series, strong on speech, education, and public-sector workflows. Limited global-market ranking.
- Tencent HY 2.0 (Hunyuan) — Tencent’s flagship language model at the time, released in December 2025 according to IT Home and TechNode, and offered through Tencent Cloud and the Yuanbao app. Minimal OpenRouter footprint: only an older open-weight Hunyuan-A13B was listed.
Note that NVIDIA's Nemotron 3 Super 120B (released March 11, 2026, 60.47% on SWE-Bench Verified with OpenHands in NVIDIA’s model card, open weights under the NVIDIA Nemotron Open Model License, up to 1M context and 262K as served on OpenRouter) often appears in "Chinese AI" conversations because of its pricing profile, but it is an NVIDIA model trained and served outside China. Do not count it toward Chinese token share.
Pricing Comparison Matrix
The table below covers thirteen flagship and volume models from eight of the ten providers, sorted by provider; Baidu and Tencent had no comparable flagship listing on OpenRouter. Figures are OpenRouter list prices and context lengths from OpenRouter’s model catalog as archived on April 3, 2026, except GLM-5, shown at Z.ai’s own API price and 200K context because OpenRouter’s default GLM-5 listing that day was capped at 80K. MiMo V2 Pro’s price applies up to 256K of context; Xiaomi charges $2 / $6 from 256K to 1M.
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Xiaomi | MiMo V2 Pro | $1.00 | $3.00 | 1.05M |
| Xiaomi | MiMo V2 Flash | $0.09 | $0.29 | 262K |
| Alibaba | Qwen 3.6 Plus | Free preview; $0.325 by April 10 | Free preview; $1.95 by April 10 | 1M |
| Alibaba | Qwen 3.5 Flash | $0.065 | $0.26 | 1M |
| Alibaba | Qwen 3 Max Thinking | $0.78 | $3.90 | 262K |
| Z.ai (Zhipu) | GLM-5 | $1.00 | $3.20 | 200K |
| Z.ai (Zhipu) | GLM-5 Turbo | $1.20 | $4.00 | 203K |
| DeepSeek | DeepSeek V3.2 | $0.26 | $0.38 | 164K |
| Moonshot | Kimi K2.5 | $0.38 | $1.72 | 262K |
| MiniMax | MiniMax M2.7 | $0.30 | $1.20 | 205K |
| MiniMax | MiniMax M2.5 | $0.12 | $0.99 | 197K |
| StepFun | Step 3.5 Flash | $0.10 | $0.30 | 262K |
| ByteDance | Seed 2.0 Lite | $0.25 | $2.00 | 262K |
For reference, OpenAI GPT-5.4 sits at $2.50 / $15.00 with 1.05M context, Claude Sonnet 4.6 at $3.00 / $15.00 with 1M, and Claude Opus 4.6 at $5.00 / $25.00 with 1M. Against those anchors, the Chinese flagship MiMo-V2-Pro undercuts US pricing by roughly 2.5-5x at input and 5-8x at output on comparable context lengths.
Capability Comparison Matrix
The capability matrix below summarizes flagship models across five workloads: coding, reasoning, tool use, multimodal inputs, and multilingual handling. The labels are our qualitative reading of the provider sections above, not independent test results; where a cell gives a figure, it is the vendor-reported benchmark or OpenRouter usage share cited there.
| Model | Coding | Reasoning | Tool use | Multimodal | Multilingual |
|---|---|---|---|---|---|
| MiMo V2 Pro (Xiaomi) | Strong (25.5% programming share) | Solid | #1 OpenRouter tool calls | Text only | Chinese + English |
| MiMo V2 Omni (Xiaomi) | Solid | Solid | Strong | Image + video + audio | Chinese + English |
| Qwen 3.6 Plus (Alibaba) | Strong (23.5% programming share) | Reasoning model | Native function calling | Text, image, video input | Broad |
| Qwen 3.5-Omni (Alibaba) | Solid | Solid | Strong | Full omnimodal | 113 languages speech |
| GLM-5 (Zhipu) | 77.8% SWE-Verified | Strong | Solid | GLM-5V-Turbo variant | Chinese + English |
| DeepSeek V3.2 | Strong | GPT-5 level (Speciale: IMO/IOI gold) | Thinking-in-Tool-Use | Text only | Chinese + English |
| Kimi K2.5 (Moonshot) | Cursor Composer 2 base | Strong | 100-agent swarm | Multimodal MoE | Chinese + English |
| MiniMax M2.7 | 56.22% SWE-Pro | Strong | Strong | Text primary | Chinese + English |
| Step 3.5 Flash (StepFun) | Solid | Solid | Solid | Text primary | Chinese + English |
| ERNIE 5.0 (Baidu) | Solid | Strong | Baidu search-native | Full omnimodal | Chinese + limited English |
Three patterns stand out. First, the coding leaders are MiMo-V2-Pro and Qwen 3.6 Plus — combined they took roughly 49% of programming-category tokens in OpenRouter’s sampled category chart for the week of March 30. Second, Qwen 3.5-Omni and ERNIE 5.0 are the most genuinely omnimodal flagships, with MiMo-V2-Omni close behind. Third, none of the benchmarks cited here measures English-language tone, so test it on your own prompts before a consumer-facing deployment aimed at US audiences.
Enterprise Readiness and Export Controls
Pricing and capability tell only part of the procurement story. The enterprise-readiness matrix below covers compliance posture, US availability, primary hosting geography, and hardware stack. These are the dimensions that typically decide whether a model passes internal review at a US or EU-based buyer.
| Provider | SOC 2 | GDPR posture | US API access | Primary geography | Hardware |
|---|---|---|---|---|---|
| Xiaomi | Not published | Limited | Direct + OpenRouter | China | Not disclosed |
| Alibaba | Alibaba Cloud SOC 2 | EU regions available | Direct + OpenRouter | China + international | Not disclosed |
| Z.ai (Zhipu) | Not published | Self-host path | Direct + OpenRouter + self-host | China | Inference on Ascend and six other Chinese chips |
| DeepSeek | Not published | Self-host path | Direct + OpenRouter + self-host | China | NVIDIA H800 (V3 base) |
| Moonshot | Not published | Self-host path | Direct + OpenRouter + self-host | China | Not disclosed |
| MiniMax | Not published | Limited | Direct + OpenRouter | China | Not disclosed |
| StepFun | Not published | Self-host path | OpenRouter + self-host | China | Not disclosed |
| ByteDance | BytePlus SOC 2 | Via BytePlus (international) | Direct + OpenRouter | China + international | Not disclosed |
| Baidu | Domestic certs | Limited | Baidu AI Cloud | China | Not disclosed |
| Tencent | Tencent Cloud SOC 2 | EU tenants available | Tencent Cloud | China + international | Not disclosed |
For regulated industries, the safer path is self-hosting open-weight models (GLM-5, Kimi K2.5, Qwen 3.5 small series) inside your own cloud region with documented data processing controls, rather than hitting Chinese-hosted APIs directly.
Models with documented support for Chinese accelerators (GLM-5 on Huawei Ascend and six other domestic chips; DeepSeek V4 was reported to target Ascend) are the procurement story for Chinese state-owned buyers under US export controls. For Western buyers, the choice is neutral to slightly negative given the smaller tooling ecosystem.
Running a model evaluation? Production selection benefits from a structured scoring process across cost, capability, compliance, and workload fit. Our Analytics Insights and CRM Automation practices translate rankings like these into measurable production wins.
Conclusion: The Q2 2026 Playbook
Chinese labs reached nearly 60% of OpenRouter tokens in early April as the leading providers converged on a consistent playbook: ship flagship models with large context windows, price aggressively below US frontier, offer free or subsidized launch access through AI coding tools, and keep a clear open-weight path for enterprise self-hosting. Xiaomi's 21.1% token share is the most dramatic data point, but the pattern runs across Alibaba, Zhipu, MiniMax, DeepSeek, and StepFun.
The actionable takeaway for Western buyers is narrower than the headlines suggest. Chinese flagship models are cheaper on list price and competitive on vendor-reported coding benchmarks and long context. They lag on published compliance posture, and English-language tone and enterprise tooling are worth testing before adoption. The sensible strategy in most production stacks is a multi-model architecture — US frontier for customer-facing English output and compliance-sensitive workflows, Chinese models for internal coding, batch processing, and cost-sensitive agent workloads, all routed through a cost-aware orchestration layer.
Translate the 2026 AI landscape into production wins
Picking the right mix of frontier and Chinese models across cost, compliance, and capability is where strategy meets engineering. We help teams route the right workload to the right provider.
Sources
Usage figures are OpenRouter token shares, not market shares. Benchmark scores are vendor-reported unless stated otherwise. Also named in the text without a link: Alibaba’s Qwen 3.6 Plus press release (April 2, 2026), ByteDance’s Seed 2.0 announcement, Kuaishou’s KwaiKAT listing, The Information’s pre-launch report on DeepSeek V4 and Huawei Ascend (as relayed by TrendForce), Reuters on Huawei’s V4 compatibility statement (April 24, 2026), IT Home and TechNode on Tencent HY 2.0 (December 2025), and the South China Morning Post on ERNIE 5.0 (January 22, 2026) and on Doubao (December 24, 2025).
Usage and benchmark rank
OpenRouter Rankings, token share by model author for the week beginning March 29 (Internet Archive snapshots bracketing the figures): April 2, 2026, 15:06 UTC · April 3, 2026, 15:20 UTC
OpenRouter Rankings, weekly model leaderboard, programming category, tool calls and the Artificial Analysis Intelligence Index as displayed: Internet Archive snapshot, April 3, 2026, 17:16 UTC
OpenRouter model catalog (list prices, context lengths, variants): April 3, 2026 snapshot · April 10, 2026 snapshot
OpenRouter rankings methodology (tokenizer and scope caveats): Datasets API documentation
Model releases, specifications and prices
Xiaomi: MiMo-V2-Pro launch page, March 18, 2026 · MiMo-V2-Flash model card
Alibaba Qwen: Qwen3.5-9B model card · Qwen3-Coder-Next model card · Qwen3.5-397B-A17B model card · Qwen3.5-Omni launch post · Qwen3.5-Omni technical report
Z.ai (Zhipu): GLM-5 technical report · GLM-5 documentation · GLM-5V-Turbo documentation · price list, April 12, 2026 snapshot
DeepSeek: V3.2 release note · V4 preview release note · API change log · DeepSeek-V3 technical report
Moonshot AI and Cursor: Kimi K2.5 announcement · Kimi K2.5 model card · Composer 2 technical report
MiniMax: M2.7 announcement · M2.7 model card and license · M2.5 model card · M2 model card
StepFun and NVIDIA: Step 3.5 Flash model card · Nemotron 3 Super model card
Compliance and later releases
Cloud compliance disclosures: Alibaba Cloud SOC 2 · BytePlus Trust Center · Tencent Cloud SOC audit
Successor models in the October 2026 notes: Z.ai release notes (GLM-5.1, April 7, 2026) · Kimi K2.6 model card · MiMo-V2.5-Pro model card · MiniMax M3 model card
Frequently Asked Questions
Deep dives on AI, marketing and development.
Practical guides and fresh insights by email. No recycled takes.
Related Guides
Continue exploring...