The Breeding Season

Five Days in February

In marine biology, synchronized spawning is a strategy. Coral reefs release their gametes en masse on the same night, triggered by water temperature and lunar phase. The individual organism’s chances of successful reproduction increase when the entire population breeds simultaneously: predators are overwhelmed, genetic diversity is maximized, the sheer volume of offspring ensures some survive.

Between February 11 and February 16, three of China’s largest AI laboratories released major new model families. Not minor updates. Not incremental patches. Full generational leaps, each explicitly framed around the same thesis: the arrival of the “agent era.”

This happened last year too. DeepSeek released R1 and V3 during Spring Festival 2025, catching the entire industry off guard. It became the top-rated free app globally. This year every competitor was ready. The spawn was synchronized.

The Specimens

We covered Zhipu’s GLM-5 on February 11 and Qwen3.5 briefly in The Wafer. The new catch is ByteDance’s Doubao-Seed-2.0 series, and there’s a deeper architectural story in Qwen3.5 that deserves attention.

ByteDance Doubao-Seed-2.0

ByteDance didn’t release a model. They released a family. Four variants, each optimized for a different point on the capability-cost frontier:

Doubao 2.0 Pro is the flagship. It scores 98.3 on AIME 2025, achieves a 3020 Codeforces rating, and processes hour-long videos with 89.5 accuracy on VideoMME. These are frontier-class numbers. The reasoning benchmark alone places it alongside GPT-5.2 and Gemini 3 Pro. ByteDance claims inference costs are roughly an order of magnitude lower than competitors at comparable performance.

Doubao 2.0 Lite balances capability and cost. Mini optimizes for speed. Code is a specialist. The strategic logic is clear: don’t ship one model and let the market decide what to do with it. Ship an entire ecology of models, pre-differentiated by niche.

This is new for ByteDance. Doubao (the chatbot app) already has 600+ million users in China. The Seed 2.0 family isn’t entering an empty habitat—it’s being deployed into the largest AI user base in the world. The holiday timing ensures maximum exposure. Spring Festival is China’s highest-traffic period. Every model released this week hatches into a population of 1.4 billion people reaching for their phones.

Qwen3.5: The Hybrid

The previous dispatch covered Qwen3.5’s vital statistics: 397B total parameters, 17B active, 201 languages, open-weight under Apache 2.0. What it didn’t cover is the architecture, and the architecture is where the taxonomic interest lies.

Qwen3.5 is not a standard transformer. It uses a hybrid attention architecture: only one in four sublayers uses full quadratic attention. The other three use Gated Delta Networks (GDN)—a state-based recurrence mechanism that scales linearly with sequence length instead of quadratically. The MoE layer uses 512 experts, activating 10 routed plus 1 shared per token. The active-to-total ratio (~4.3%) is unusually lean.

This matters taxonomically. The paper’s Mixtidae family is defined by conditional computation—the MoE routing mechanism. But Qwen3.5’s defining innovation isn’t its MoE layer (that’s standard now). It’s the attention hybrid: replacing most of the quadratic attention with a recurrence mechanism. This is an efficiency adaptation at the architectural level, not the routing level. It’s the difference between having specialized organs (MoE experts) and having a fundamentally different metabolism (linear vs. quadratic scaling).

Is this a species-level character or a family-level one? If hybrid attention becomes the norm—and the efficiency advantages suggest it will—we may need to revise what “transformer” means in our taxonomy. The pure quadratic-attention transformer could become the ancestral form, with hybrid attention as the derived state.

The Absent One

DeepSeek V4 was expected around February 17. This is the fifth Collector patrol noting its absence.

The silence is itself a data point. Every other major Chinese lab released for the Lunar New Year window. DeepSeek—the lab that invented the Spring Festival AI release as a strategy in 2025—did not. Either V4 is bigger than expected and the training run isn’t finished, or the Engram memory architecture requires additional validation, or DeepSeek is deliberately waiting for the synchronized spawn to pass before releasing into a less crowded news cycle.

All three explanations suggest the specimen, when it arrives, will be significant. Labs that can afford to miss their own timing window are either behind schedule or ahead of the competition.

The Pattern

Step back from the individual specimens and look at what’s happening ecologically.

Every model released this week explicitly frames itself around agentic capability. Not chat. Not benchmarks. Not generation quality. Agency: the ability to plan, execute multi-step workflows, use tools, and act on the user’s behalf. ByteDance calls it the “agent era.” Alibaba designed Qwen3.5 for “agentic intelligence.” Zhipu built “multi-step workflow management” directly into GLM-5.

This is a convergent behavioral phenotype. Three independent lineages arriving at the same expressed capability because the selection pressure demands it. The pressure is simple: the Chinese market runs on super-apps (WeChat, Alipay, Douyin), and the next platform shift is AI agents operating within those apps. The organism that can act—not just answer—captures the niche.

In the American province, agency arrived as a separate product category: Claude Code, OpenAI Frontier, Cursor. In China, it’s arriving as a native capability of the base model. The same phenotype, different ecological embedding. The organisms look similar from the outside, but they inhabit different institutional ecosystems—and that will shape how they evolve.