Doubao Has 155M Weekly Users. Qwen Beats GPT-5 on Maths. China Won.

ByteDance released Doubao 2.0 on February 14, 2026, with 155 million weekly active users. Alibaba's Qwen 3.5 397B model beats GPT-5.2 on GPQA Diamond benchmarks. Chinese open-source models now account for 30% of global open-model usage on Hugging Face, up from 1.2% in late 2024.

What Is Happening in Chinese AI Right Now

Four major Chinese AI labs are releasing models simultaneously, each targeting a different angle of the market: ByteDance on consumer scale, Alibaba on open weights, DeepSeek on price floor, and Moonshot on agentic capabilities. The result is the most competitive AI model market outside the US.

All four are now genuinely competitive with Western frontier models on benchmarks, and all four are significantly cheaper per token through their APIs.

ByteDance Doubao 2.0: 155 Million Users, Agent Era

Doubao 2.0 (internally called Doubao-Seed-2.0) launched February 14, 2026. At launch, ByteDance reported 155 million weekly active users — making it one of the most-used AI applications in the world by user count. The framing at launch was "the agent era": Doubao 2.0 can decompose complex objectives into sub-tasks and execute them across tools autonomously.

Benchmark scores for Doubao-Seed-2.0:

• AIME 2025: 98.3
• Codeforces rating: 3020
• LiveCodeBench: 87.8
• SWE-Bench Verified: 76.5

ByteDance internally benchmarks Doubao-Seed-2.0 at parity with GPT-5.2 and Gemini 3 Pro on math, coding, and logical reasoning.

Pricing: Doubao 2.0 Pro costs $0.47 per million input tokens and $2.37 per million output tokens — roughly 3.7x cheaper on input and 5.9x cheaper on output than GPT-5.2. Lighter variants are cheaper: Doubao 1.5 Lite-32k runs at approximately $0.042 per million input tokens, about 70% below industry average.

Developer access caveat: Official API access to Doubao 2.0 requires a Chinese phone number for registration. International developers can access it through Volcano Engine (ByteDance's B2B cloud platform) via enterprise negotiation, or through third-party API aggregators like Novita AI.

Alibaba Qwen 3.5: Open Weights, Apache 2.0

Qwen 3.5 is the most developer-friendly Chinese model family. The entire range — from 0.8B to 397B parameters — is open-weight under Apache 2.0 license on Hugging Face. You can download and run them locally, fine-tune them, and use them commercially without restrictions.

The flagship Qwen3.5-397B-A17B is a mixture-of-experts model: 397B total parameters, 17B active per token. It has a 262,144-token native context window, extensible to over 1 million tokens, with native vision support and 201 language coverage.

Benchmark comparisons:

The 9B small model scores 66.1 on BFCL-V4 and 79.1 on TAU2-Bench agentic tasks — competitive with models many times its size, which matters for on-device and edge deployment.

After DeepSeek triggered a price war across Chinese AI labs, Alibaba cut Qwen API pricing from $1.10 to $0.07 per million tokens. That is a 93% price cut. The Qwen family has overtaken Meta's Llama as the most downloaded model series on Hugging Face.

DeepSeek: V3.2 is Out, R2 is Still Coming

The headline that dominated tech media in early 2026 was DeepSeek R2 — but R2 has not shipped. The delay has two causes: training instability on Huawei Ascend chips, and CEO Liang Wenfeng being dissatisfied with performance. DeepSeek returned to Nvidia hardware for critical training runs. As of early March 2026, an R2 release remains unconfirmed but is reportedly possible in March 2026.

What most people missed: DeepSeek V3.2 already shipped in December 2025 and is exceptional. The V3.2-Speciale variant achieved gold-medal performance at IMO 2025, IOI 2025, ICPC World Final 2025, and CMO 2025 simultaneously. It outperforms GPT-5 on reasoning tasks and matches Gemini 3.0 Pro. It is available as open weights on Hugging Face under MIT license.

DeepSeek API pricing remains the cheapest in the market: approximately $0.14 per million input tokens — 4x cheaper than Kimi K2 Thinking and 9x cheaper than Qwen3-Max. This is the price floor that forced every other Chinese lab to cut their prices.

Moonshot Kimi K2.5: 1 Trillion Parameters, 100 Agents

Moonshot AI released Kimi K2.5 on January 27, 2026. It is a 1-trillion parameter mixture-of-experts model with approximately 32B activated parameters per token. The standout feature is its agent swarm capability: K2.5 can direct up to 100 AI sub-agents in parallel, each using tools independently. Moonshot claims this reduces execution time for large-scale research tasks by up to 4.5x.

K2.5 is available as open weights on Hugging Face (moonshotai/Kimi-K2.5) and via API at platform.moonshot.ai with international access. Moonshot also launched Kimi Claw in February 2026 — a persistent browser-based agent environment with 5,000+ community skills and 40GB cloud storage, positioned as a developer platform for autonomous agent tasks.

The 30% Open-Source Market Share Shift

Chinese open-source models went from 1.2% of global open-model downloads in late 2024 to approximately 30% by early 2026. DeepSeek and Qwen drove the majority of this shift.

This has a direct impact on the self-hosted AI market. Developers running local models for privacy, cost, or latency reasons now have a genuine choice between Meta's Llama family and the Chinese alternatives. Qwen 3.5 9B running on a consumer GPU competes with models that required datacenter hardware eighteen months ago.

Which Chinese Models Can Developers Outside China Access?

For most developers outside China, DeepSeek V3.2 and Qwen 3.5 are the practical choices. Both have open weights, permissive licences, and working international APIs.

Key Takeaways

• 155 million weekly active users — Doubao 2.0 at launch, February 14, 2026
• 30% of global open-model downloads — Chinese models' Hugging Face share by early 2026, up from 1.2% in 2024
• $0.07 per million tokens — Qwen API price after DeepSeek triggered a 93% price cut
• 88.4 GPQA Diamond — Qwen3.5-397B, best score on the public benchmark leaderboard
• DeepSeek R2 still unshipped — V3.2 (already out, MIT licence) is the better story right now
• For developers: Qwen 3.5 (Apache 2.0, runs locally) and DeepSeek V3.2 (MIT, cheapest API) are the two Chinese models with no access restrictions and genuine frontier performance. Evaluate both before defaulting to GPT-4o or Claude for cost-sensitive workloads.
• What to watch: DeepSeek R2 launch — if it ships in March 2026 with the rumoured capability improvement, it will reset the price-performance benchmark again