Meituan Open-Sources 1.6-Trillion-Parameter Agent Model LongCat-2.5, Trained Entirely Without Nvidia Chips

• Dispatched: Nov 15, 2026 · 08:38 ET (13:38 UTC)

One-sentence summary: Meituan has released the public weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model for AI agents trained entirely on Chinese hardware with no Nvidia GPUs involved, scoring close to Claude Opus 5 on benchmarks and released under the MIT license.

Verification status: ✓ Primary disclosure (model weights and self-hosting guide published on Hugging Face and GitHub, reported by The Decoder); benchmark scores are vendor-reported and not yet independently verified

Core Facts & Consensus Delta:

Meituan's LongCat team released the weights for LongCat-2.5 on 2026-11-15 05:06 ET / 10:06 UTC, about one week after the model's official unveiling. LongCat-2.5 is a mixture-of-experts architecture with 1.6 trillion total parameters, activating roughly 48 billion per token, built for long-horizon, multi-step autonomous agent tasks — especially coding — with text, image and video input, up to one million tokens of context via the API, responses up to 128,000 tokens, and a switchable reasoning mode.

The hardware path is the point of the release: the model was trained entirely on Chinese hardware, with no Nvidia GPUs used. Meituan quantizes expert weights and activations to 4-bit precision, which Chinese chips process natively; the self-hosting guide describes a sample deployment of 194 Huawei Ascend Atlas A2 accelerators (two for image processing, 64 for prompt ingestion, 128 for response generation), with no guidance for Nvidia GPUs. Meituan says the 4-bit version delivers 43.3% higher prompt-processing throughput and 42.77% higher generation throughput than the 8-bit version, with accuracy on par with the uncompressed model. Its predecessor, LongCat-2.0, was trained in late June 2026 on a cluster of more than 50,000 Chinese AI chips.

On benchmarks (all vendor-reported, not yet independently verified): Terminal Bench v2.1 scores 88.4, close to Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, below DeepSeek-V4.1-Flash at 90.6; DeepSWE scores 64.6, up sharply from the predecessor's 9.7; the harder Terminal Bench v4.0 scores 38.5, well below Opus 5's 53.9. The weights carry the MIT license, permitting commercial use, modification and redistribution; the API supports OpenAI and Anthropic formats and plugs into agent harnesses such as Claude Code, Codex and Hermes, with a limited-time promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens.

Supply-Chain & Market Read-Through: This is the first open-weights model at this scale shown to be trained and served entirely without Nvidia hardware, and it matters on three transmission paths. First, it tests the real-world effect of US export controls: blocked chips do not equal blocked frontier-model training, since domestic accelerators plus 4-bit quantization can now support a near-top-tier agent model. Second, it gives the Huawei Ascend ecosystem a verifiable use case; if more Chinese labs follow the same route, demand for domestic accelerators and their supporting memory will rise. Third, for Nvidia, hardware substitution from Chinese suppliers is moving from rhetoric to downloadable, deployable open weights, eroding its hardware moat in the Chinese market. Caveats apply: training cost and cluster size for LongCat-2.5 were not disclosed, benchmark scores are vendor-reported and await independent verification, so this is not yet evidence that Chinese hardware has fully caught up with Nvidia. For HBM and memory makers (SK Hynix KRX: 000660, Samsung Electronics KRX: 005930, Micron MU, SanDisk SNDK), the release does not directly change memory supply and demand: training ran on Chinese chips and their own memory stack, so any demand uplift accrues first inside the domestic Chinese supply chain.

Next Catalysts & Watchpoints:

• Independent third-party benchmark replication (e.g., LMArena or independent evaluators) to confirm the reported 88.4 Terminal Bench score.
• Huawei Ascend shipment and adoption signals, and whether more Chinese labs disclose non-Nvidia training routes.
• Any US export-control policy response; Meituan API adoption data and subsequent model updates.

Sources:

• The Decoder (Published: 2026-11-15 05:06 ET / 10:06 UTC)
• Agibytes (Published: 2026-11-15, after 10:06 UTC)
• Oossa (Published: 2026-11-15 05:22 ET / 10:22 UTC)

Information only, not investment advice.