Skip to content

Latest commit

 

History

History
98 lines (67 loc) · 3.37 KB

File metadata and controls

98 lines (67 loc) · 3.37 KB

Upstream demos queued for reproduction

These examples answer the useful question: what does the model actually make or do? They are original third-party demos, not Y outputs. Hardware and performance statements remain source-reported until a separate Y proof file reproduces them.

Coding agent

Qwen3-Coder-Next

The Qwen Team's official use cases show the agent building and deploying a website, creating games, cleaning a desktop, testing a website, and producing an audiovisual coding project.

Small local research agent

Qwen3.5-4B via Unsloth

Unsloth's X video shows a 4B model searching more than 20 sites, calling tools, and returning citations. Unsloth reports a 4GB RAM footprint; Y has not yet verified that claim.

Computer-use agent

Microsoft Fara1.5

Microsoft Research's official trajectories show screenshot-driven agents shopping, completing forms, booking, researching, and pausing for approval.

Actual phone inference

Google AI Edge Gallery

Google's open-source phone app demonstrates offline chat, image understanding, on-device transcription, mobile actions, agent skills, and device-specific benchmarking.

Image generation

Z-Image

Tongyi-MAI's official gallery shows photorealism, English and Chinese text rendering, and prompt adherence. Its released Turbo model is reported to fit within 16GB VRAM.

Video generation

Wan2.2

The official Wan Team examples show cinematic text-to-video and image-to-video with complex motion. The project reports that TI2V-5B can produce 720p at 24fps on an RTX 4090-class 24GB GPU.

Full-duplex voice

NVIDIA PersonaPlex

NVIDIA Research's audio examples demonstrate interruptions, contextual backchannels, persona control, and instruction following across assistant, banking, medical-intake, and emergency scenarios.

Reproduction rule

We do not copy an upstream result into a Y product claim. To become Y verified, a demo must be rerun with a pinned model, quant, runtime, commit, prompt, and exact target device. We publish the output and the failure modes, not only the best number.