Unisound released the next-generation general large model U2 on June 8, slapping the "native Agent large model" label on itself. U2 has a total of 266B parameters, activating only about 10B per inference — compared to dense models that easily exceed a trillion, the token consumption is about 25% of that, and "intelligence density" is embodied in this set of numbers.
On the technical path, U2 takes a route different from most peers: internalizing tool calling, state management, and task planning into the model from training day one, rather than "training a chat model first and then externalizing an Agent framework." For this, Unisound introduces a model + Harness co-evolution mechanism — the more complex the model body becomes in training, the more the support nodes and validation precision of the controlling scaffold also extend in sync; the more rigorous the scaffold, the more it constrains the model output in return, forming a closed loop.
There are also two key training-side designs: a process supervision + curriculum learning combination lets the model evolve gradually from easy to hard, from short-context to long-chain tasks; implicit thinking reasoning + explicit thinking verification — daily exploration stays in latent space, avoiding the token waste of explicit reasoning, and switches to readable, verifiable explicit reasoning at key stages. This is the opposite of the "full-chain-of-thought" path of the o-series and R1.
The benchmark performance also holds up. GPQA Diamond 87.9, beating GLM-5.1 (86.2), DeepSeek-V4-Flash High (87.4), and MiniMax M2.5/M2.7 (87.4); IFBench instruction following 77.3; AA-LCR long context 70, beating GLM-5.1 (62.3) and DeepSeek-V4-Flash (62.7). Claw-Eval pass@3 takes 76.9, putting Agent capability at the front of the domestic first tier.
Unisound is turning "high intelligence density × high token value" into a product formula, essentially answering: when compute is no longer infinite, what enterprises really buy is "how much reliable task completion per dollar of token." U2's 266B/10B MoE choice, Harness co-evolution, and implicit/explicit reasoning switching engineer this path. "Few parameters can still run a deliverable Agent" is quietly rewriting the value-evaluation standard for domestic foundation models.