Meituan's LongCat-2.5, Hume's Audio Decision API, and More

The biggest model news on Monday came from Meituan. LongCat-2.5 is a 1.6T-parameter model with about 48B active parameters, natively multimodal, and built for long-horizon agent work like visual coding, creative design, and 3D tasks.

Meituan says it was trained on Chinese chips and places it between Opus 4.8 and Opus 5 on its own benchmarks. Independent checks are still pending, so I'd treat the ranking as a claim for now. The part I find more interesting is the training hardware: a frontier-adjacent agentic model trained outside the usual Nvidia stack is a real signal, if the numbers hold up. (Meituan's announcement, LongCat blog)

What Else Happened

Hume launched the Audio Decision API. You ask plain-language questions of raw audio, like how many speakers there are, whether anyone is laughing, or whether the caller sounds frustrated, and you get probability answers back with no transcript step. Hume says it replaces 10+ models, was trained on 10M+ human judgments, and hits 70.7% agreement with human panels. Access is by waitlist for now. (Hume's announcement)

Deep Principle released Hydrogen-1.0, an open-weight science model. It's GLM-5.3 post-trained on internal lab records and discovery trajectories, and it reports 38.3% on FrontierScience-Research, against GPT-6 Astra's 33.3%, plus 93.0% on GPQA-Diamond. Training on how real discoveries actually unfolded, not just on papers, is a smart angle. (Chenru Duan's announcement)

OpenAI introduced LifeVal and pushed harder on security. LifeVal measures everyday personal-assistant tasks like childcare, government services, and personal finance, shaped by feedback from thousands of ChatGPT users. Separately, OpenAI is partnering with Cathedral Systems on its next internal security sprint and more than quadrupling critical bug bounty rewards, up to $172K. (LifeVal, security sprint)

OpenAI also built Pando, a new Git backend for its monorepo. It cut push latency by about 80% and serves nearly a petabyte of Git data a day, and it was built to scale with agent usage. When agents become the main committers, the source-control layer is one of the first things to buckle. (OpenAI Devs post)

Addy Osmani wrote about "cognitive surrender" when agents do the work. It's a long, practical essay on keeping your own judgment while running agents: form a hypothesis before you ask, name what every green check didn't check, and read what the agent actually ran before reading its summary. This is the one I'd send to anyone on a team leaning hard on agents. (Addy's article, related blog post)

Affinity for iPad is live, and the core is free. Canva shipped the full Affinity suite, with sketching, vector, photo editing, and page layout, on iPad with Apple Pencil support. AI features like generative expand need a Canva premium plan. Designers are calling it desktop power rather than a companion app. (Affinity's announcement)

Launches and Demos

DeepSeek Harness Creator Mode – describe a plugin in plain English and the harness writes, installs, and runs it, with demos including a design workspace, a sandbox game, and a 3D LLM visualizer.
Figma Shader Town update – render and export in 4K, and export straight into Code Layers, rolling out this week.
Framer Agents Harness 2 – agents start editing while still working through your prompt, and Framer says it's up to 33% cheaper and 20% faster on long tasks.
Workers IO (YC F26) – runs millions of simulated production failures like network delays, disk errors, and clock skew inside a deterministic hypervisor, so every failure replays exactly. It's pitched as verification for agent-written code.
Entelligence Memory – one shared memory graph of sessions, PRs, tickets, and incidents across Claude Code, Codex, and Cursor, claiming 53% fewer repeat errors.

Accounts Worth Following

@chaaai – Chaitanya, founder of Workers IO, writing about software reliability and verifying agent-written code.
@chenru_duan – Chenru Duan, Deep Principle's founder and CTO, with a front-row view of open-weight AI for science.
@avstorm – Andreas Storm, designer and iconographer at Iconists, sharp on design tools and craft. He flagged the Affinity iPad launch early.

That's it for today. 👋