GeekPark Hands-On, Deep Read: CPO Report, 3D Watch and the NAIAD Remake

GeekPark published the most complete first hands-on test of Hy4 in the Chinese press: three lines — a long-horizon investment-research task, visual code generation, and a game remake — each producing a verifiable deliverable rather than just a benchmark score. This article reads that test closely and extracts its real judgment of Hy4's capabilities.

Line one: the CPO research task, a "junior research assistant" prototype

GeekPark assigned Hy4 a CPO (co-packaged optics) investment-research task, asking for a research report and a valuation model. Hy4 delivered a report PDF and a recalculable valuation spreadsheet (xls), assessed as a "junior research assistant" prototype — able to complete the full loop of gathering material, structuring it and building the numbers, but with limited depth of research judgment. That assessment points the same way as the official FiT finance crowdsourced test (3.04, best among open models): finance analysis is a relative strength for Hy4, but there is still distance between "assistant" and "analyst."

Line two: the Three.js 3D mechanical watch, full interaction loop

Line two was visual code generation: build a Three.js 3D mechanical-watch explainer page. Hy4's delivered page ran its full interaction loop — exploder animation, part highlighting, camera controls forming a complete interaction cycle. This combines vision understanding with code generation: the model first has to "see" the watch's structure, then write it as an interactive 3D page. GeekPark's judgment: vision makes the product more useful, but doesn't decide the race — because vision is becoming standard across flagship models.

Line three: the NAIAD game remake, a visual self-check loop

The most technically interesting line was remaking the web version of the game NAIAD. Hy4's workflow was "watch frames → rebuild → render and self-check": first watch an 81-second gameplay recording, extract frames to understand the gameplay and visual style, rebuild it in Canvas 2D, then run 4–5 rounds of self-checking through a headless browser — render it, screenshot it back, spot problems, fix them. That "see–build–verify" loop is a combined demonstration of Hy4's vision and computer-use capabilities, and a task form the preview simply couldn't attempt.

The test's conclusion: a threshold, not a destination

GeekPark's overall judgment: Hy4 has reached the open-camp first-tier threshold, but is still chasing the global top; the real test will be Hy5. That matches the official data: AA index 35, No. 2 in China, tying Grok 4.7; Terminal-Bench 4.0 44.4%, behind Kimi's 66.8% and Claude's 66.4%; but best among open models in the finance crowdsourced test, and only 2.9 points behind Claude on OSWorld 2.1.

Notably, GeekPark's test didn't touch Hy4's two most contested points: the 89% hallucination rate on AA-Omniscience (independent measurement) and the unpublished full-release weights. The sample is also small — three lines, one run each — and can't represent Hy4's stable performance. InfoQ's WorkBuddy test fills in the other side: a tower-defense game completed in three rounds over 55 minutes, but with late-game playability dropping; document reconciliation with a counting error and an unfounded inference. Read together, the two tests paint Hy4 as: real change in task-closure capability, details still needing human sign-off.