← All episodes
Daily AI Brief

Meituan's 1.6-trillion-parameter open model never touched an Nvidia GPU

Sunday, November 15, 2026 · 29 stories · 11:35 · 1326 words
Meituan's 1.6-trillion-parameter open model never touched an Nvidia GPU

Watch on YouTube · subscribe for the daily brief

The LongCat team at Meituan, the Chinese food delivery company, released the weights for LongCat-2.5, a model built for agents that work through long coding tasks on their own, The Decoder reports. It was trained entirely on Chinese hardware, and the license is MIT, which allows commercial use.

What it is

  • Size: a mixture-of-experts model with 1.6 trillion parameters, about 48 billion active per token. Through the API it takes up to one million tokens in and writes up to 128,000 out. It reads text, images and video.
  • Hardware: the self-hosting guide describes 194 Huawei Ascend accelerators (Atlas A2): 2 for images, 64 for prompt ingestion, 128 for generation. There is no guidance for Nvidia GPUs. The predecessor LongCat-2.0 was trained on more than 50,000 Chinese AI chips.
  • 4-bit by design: experts' weights and activations are quantized to 4-bit, which Chinese chips process natively. Meituan reports 43.3% higher prompt throughput and 42.77% higher generation throughput than 8-bit, with accuracy on par.
  • No shortcuts in training: for reinforcement learning on real GitHub pull requests, the team blocked internet access, removed Git history and cached files, and had a model check solutions for tricks.

How it scores (Meituan's own table)

  • Terminal Bench v2.1: 88.4, against Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and DeepSeek-V4.1-Flash on top at 90.6.
  • Terminal Bench v4.0: 38.5, just above GPT-5.6 Sol's 37.3 and well below Opus 5's 53.9.
  • DeepSWE: 64.6, up from 9.7, behind every model in the table except Claude Opus 4.8. DeepSeek-V4.1-Flash leads at 74.2.
  • BrowseComp: 91.9, the top score in the table, run in Meituan's own environment.

The newest US models, Claude Opus 5.5 and GPT-6.1 Sol, are not in the table. The level matches what US models reached a few months ago.

What to take from it: the API speaks both the OpenAI and Anthropic formats and plugs into harnesses like Claude Code, Codex and Hermes, so trying it costs one config change. Before you run any open weights with tools, put the same egress rules on them as on any third-party binary. An open model tuned for Huawei chips from day one also makes that hardware more attractive outside China, which is the part export policy cannot see in a benchmark table.

Nvidia: $12 billion through Megaspeed, and Congress moves

  • The orders: Megaspeed arranged nearly $12 billion in Nvidia high-performance chip orders for Chinese companies between March 2024 and June 2025, the Wall Street Journal reported from documents, per The Decoder. Nvidia officially reported $12.4 billion in China revenue over six months, 7% of the total. Short seller Culper Research estimates more than 20% of last fiscal year's revenue came from Chinese customers.
  • Megaspeed's line: it says Alice Huang left on August 30, 2025, has no current role, and that its business was "legitimate and lawful." Spy allegations against her are not confirmed.
  • Congress: both parties want tighter chip export rules. Chuck Schumer says Huang and the administration must answer to Congress, Senate Banking chair Tim Scott asked Huang and Howard Lutnick to testify, and the stalled Remote Access Security Act is back in play. Speaker Mike Johnson cites "growing consensus in Congress on limiting chip exports to China," per Reuters.
  • A comeback? The Financial Times reports Huang has discussed returning as CEO with close allies, The Verge notes. The reinstatement letter passed 150,000 signatures, but Business Insider found names like "Tim Cruise, Hollywood actor."
  • The cost so far: Nvidia closed at a record $239.24 on October 6 and at $162.85 on November 13, The Verge's timeline of the scandal shows. Harvey Jones became the first board chair in company history. Nvidia has invested $99 billion in AI labs, startups and neoclouds, which is why TechCrunch's Equity crew asks whether his fall hurts the whole industry: there was no succession plan.

AI and jobs: young workers first

Australia's central bank built a monitor for the places AI effects should show first. Job postings fell across all groups since late 2022, steepest in the most exposed jobs like programming, clerical work and customer service. There, workers aged 15 to 24 and young graduates have weaker employment and higher unemployment than older workers.

  • The caveat: the analysis, by Mike Major, does not establish cause and measures theoretical exposure, not actual use. Business cycles and remote work could drive the numbers.
  • Same pattern elsewhere: a 2025 Stanford payroll study found employment of 22- to 25-year-olds in exposed jobs down 13% in relative terms, nearly 20% for young software developers. A Denmark study of 25,000 workers found no change in wages or hours, with 2.8% of time saved.
  • Against blaming AI alone: a University of Pittsburgh study of 10.6 million LinkedIn profiles found the trouble in exposed jobs began in early 2022, before ChatGPT.

Coding models: same score, different bill

The New Stack ran Claude Fable 5.2 against Opus 5.5 on three coding and agent jobs at API defaults (Fable on High, Opus on Medium), each made harder twice. Both scored 50/50 on all nine runs. Fable cost almost four times as much and took more than twice as long: on the cron scheduler, $3.29 and 661 seconds vs. $0.99 and 358 seconds. Fable's list rates are twice Opus's. A day earlier, the same outlet found Fable used five times the tokens of GPT-6 Astra for tied scores.

The pattern: the model at the top of the leaderboard is rarely the right default. Start with the cheaper setting and move up only where your own tests fail.

Agents and security

  • Uber's Secure AI framework checks whether an agent's action fits the request, not only whether it has permission. Agents register in an AI Registry with owner, persona, permitted intents and constraints; a tool called Argus scans their code, prompts and tool paths; runtime hooks call AI Guards before actions; and recorded sessions are replayed when policies change. Uber rates explicit per-agent intents and constraints as the most effective control, and generic thresholds as a failure.
  • Apple's health AI study: Apple and OpenEvidence launched a study app for up to 200,000 US participants over 21, for more than 12 months. Answers cite sources like NEJM and JAMA, users can share HealthKit data, data is anonymized and not used for training, and OpenEvidence deletes queries after 24 hours. The app is not tied to Siri, The Verge reports.
  • Bharat Semi raised a $24 million seed for image sensors that capture, process and run AI on one chip, which it says cuts system cost by at least 70%. It is pre-revenue, with design wins at General Atomics and Kaynes Semicon, and targets production in 2027.

Data tools

  • DuckDB's Quack is an HTTP-based protocol that streams query results in column batches as DuckDB produces them, so you can work with results while the query runs. Older protocols send row by row and repeat column names and types, "quite a bloated message," says DuckLabs' Guillermo Sanchez Dionis. It shipped in DuckDB 2.0.0 LTS in October.
  • Datasette 1.0a43 adds streaming for output renderers (CSV and TSV exports use it) and CSS multi-page view transitions. It fixes a bug that could expose the title and SQL of private stored queries. Datasette 0.65.6 fixes a database download that bypassed access rules when its route differed from its name: upgrade if you use `add_database(..., route=...)`.
  • More from Simon Willison: datasette-maplibre maps data with MapLibre GL and OpenFreeMap, datasette-apps 0.4a0 lets sandboxed apps link to each other by URL fragment, and datasette-agent 0.6a0 puts an agent prompt on the homepage.
  • Sebastian Raschka: Understanding positional embeddings walks from learned absolute positions to RoPE, with a small PyTorch implementation and the choices made in 2026 open-weight models.

The through-line

Chips, models and money are moving around Nvidia, not through it: an open frontier-class model trained on Huawei hardware, $12 billion in orders under investigation, and Congress writing new rules. Today, check which of your AI dependencies assume one chip vendor, one model, or one price, and note what you would switch to.

Sources covered

1 story · 29 sources (20 news, 3 company, 6 voices)

  1. datasette-maplibre 0.1a0
    🎙️Simon Willison↗🎙️Simon Willison↗🎙️Simon Willison↗🏢OpenAI↗🏢OpenAI↗📰TechCrunch↗📰The Verge↗🏢Vercel AI↗📰TechCrunch↗📰Engadget↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰InfoQ↗📰The Verge↗📰BleepingComputer↗🎙️Sebastian Raschka↗📰The Hackers News↗📰The Decoder↗📰BleepingComputer↗📰The New Stack↗📰The New Stack↗📰The Verge↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰InfoQ↗🎙️Simon Willison↗🎙️Simon Willison↗