AI Prompt
BlogAbout
← Back

Tencent's Hy4, Anthropic's FlashProve, and More

4 min read

The biggest model launch on Tuesday came from Tencent. Hy4 is a 770B-parameter mixture-of-experts model with about 49B active parameters and a 1M-token context, and it can both see images and operate computers across command-line and GUI apps.

It's live on Tencent Cloud and OpenRouter now, with full weights and a lighter Flash version coming later. Computer use has mostly been a closed-model feature so far, so an open model that ships with it built in is worth watching once the weights actually land. (Tencent Hy announcement)

What Else Happened

Anthropic introduced FlashProve for provable inference. It uses established cryptographic proofs from the zkSNARK family so developers can verify that a given response really came from a given model, and Anthropic says it's far more efficient than earlier open-source methods. The code is being open-sourced, which matters more to me than the announcement itself, since verification only helps if anyone can run it. (Anthropic on FlashProve)

Designers are calling Opus 5.5 the first model with "taste." shadcn posted "We've been waiting for this. Opus 5.5 is the first model with taste," then followed up that taste only shows after weeks of prompting and product building, not on a next-day benchmark. I agree with the second half: design quality is the kind of thing you notice over a project, not in a leaderboard. (shadcn's post, his follow-up, X story roundup)

Revolut moved 2,000 engineers to self-hosted open-weight coding models. It runs GLM 5.3 on 56 Nvidia B200s at about 30 cents per million tokens, roughly a tenth of its earlier cost, and reports 60% faster coding time since April. It still keeps Claude and Codex for harder jobs, which feels like the realistic split for most large teams. (X story roundup)

Intology's Locus automates molecular-design AI research. Locus proposes experiments, runs them on GPUs, scores the results, and iterates to build scientific models. On adapting Boltz-2 for potency prediction at equal compute, it reportedly beat Codex and Claude Code, with a Spearman score of 0.622 versus 0.569 and 0.450. (Intology announcement)

Sarvam released Prism, a decision model that never writes text. Built on Sarvam 105B, it picks from given options in a median 44ms with a score of 71.8, and it's aimed at routing, triage, and moderation. The weights are open, and a model that only chooses is a neat fit for the boring plumbing inside agent systems. (Sarvam announcement)

Figma Code Layers is now available to everyone. It's out of preview, and you can wire up complex flows with prototype noodles. Figma also teased another launch for today. (Code Layers for everyone, prototype noodles, the tease)

Pruna AI open-sourced P-H3-Image. It's an image generation and editing model made from MiniMax's H3 video model, and it runs in 8 steps instead of 50, about 5x faster. Weights are on Hugging Face. (Pruna AI announcement)

Sentient shipped EvoSkill v2. It improves agent skills on its own by testing the agent, learning from failures, and rewriting its instructions, without changing the model. It's open source. (Sentient announcement)

Two free stealth models showed up on OpenRouter and OpenCode. Crouton Alpha and Pulse Alpha both offer 1M context and multimodal input in a limited preview, and who made them hasn't been disclosed. (X story roundup)

Design and Building

  • Wanaka Studio 1.0 – an AI-native game engine for building 3D multiplayer games with no code and publishing them as browser links; its speed and cost claims come from the company's own tests.
  • Pixelspec – a new tool for designing on the web alongside coding agents, from a designer who credits Agentation as the inspiration.
  • Figma canvas plus MCP as agent memory – a designer holding a 3D design system, sample UI, specs, and Code Layers in Figma for Three.js web game builds.
  • Sumiro – a browser-playable 3D ink-wash game built in 24 hours with Flux, Tripo, and Claude in Three.js, which sparked a debate over prototypes versus real game dev.

Accounts Worth Following

  • socrates (@sochar__) – a designer, cofounder of quint_sft, and Friends of Figma Athens lead building agent-ready 3D design systems with Code Layers and MCP.
  • Intology (@intology) – a lab automating scientific discovery; Locus is one of the clearest examples of agents doing real research work.
  • Yesith Thomas (@yesith_thomas) – a design and marketing builder shipping Pixelspec.

That's it for today. 👋