Log in Sign up
News AMD's August Deal for Inference Chip Startup Taalas Was Worth $2.9 Billion, Filing Shows, With $1.4 Billion Contingent Business Google Expands Voluntary Exit Buyouts to Five More Countries, With Swiss Staff Offered Up to 12 Months' Pay Apple Intelligence Apple Cuts Engineers' Daily AI Spending Cap to $200, With Deeper Cuts on the Table Meta Meta Buys Conductor, the Startup Whose App Runs Multiple AI Coding Agents in Parallel News MoWorld 2.0 Debuts at Wuzhen: a 4D World Model Trained Entirely on Huawei's Ascend Security Sentra Raises $150 Million From Capital One Ventures to Rein In What AI Agents Can Read Business Kuaishou's Chip Spinoff Lingchuan Raises 1 Billion Yuan to Cut the Cost of AI Video Productivity Y Combinator President Garry Tan Says He Spends Up to $200,000 a Year on AI Tokens News AMD's August Deal for Inference Chip Startup Taalas Was Worth $2.9 Billion, Filing Shows, With $1.4 Billion Contingent Business Google Expands Voluntary Exit Buyouts to Five More Countries, With Swiss Staff Offered Up to 12 Months' Pay Apple Intelligence Apple Cuts Engineers' Daily AI Spending Cap to $200, With Deeper Cuts on the Table Meta Meta Buys Conductor, the Startup Whose App Runs Multiple AI Coding Agents in Parallel News MoWorld 2.0 Debuts at Wuzhen: a 4D World Model Trained Entirely on Huawei's Ascend Security Sentra Raises $150 Million From Capital One Ventures to Rein In What AI Agents Can Read Business Kuaishou's Chip Spinoff Lingchuan Raises 1 Billion Yuan to Cut the Cost of AI Video Productivity Y Combinator President Garry Tan Says He Spends Up to $200,000 a Year on AI Tokens

Meituan Ships LongCat-2.5: 1.6 Trillion Parameters Trained End to End on Domestic Chips, From 2 Yuan per Million Tokens

Chatbots · By: AIReport Admin · 16 hours ago · 👁 18
Meituan Ships LongCat-2.5: 1.6 Trillion Parameters Trained End to End on Domestic Chips, From 2 Yuan per Million Tokens

Meituan released LongCat-2.5, a 1.6-trillion-parameter MoE that activates 48 billion weights per token, with 1M-token context and native vision, trained entirely on Chinese-made accelerators. The company says single-card training throughput rose 86 percent over a DDP baseline and W4A4 inference lifted throughput about 40 percent, figures not independently verified.

Meituan released LongCat-2.5 on November 2, the second generation of its in-house large model, and the numbers that matter most are not on any benchmark. The model carries 1.6 trillion total parameters and activates about 48 billion per token, was trained end to end, pretraining through inference optimization, on Chinese-made AI accelerators, and is served with a 1-million-token context window and up to 128,000 output tokens.

Pricing undercuts the field: 2 yuan per million input tokens (0.04 yuan on a cache hit) and 8 yuan per million output tokens, and Meituan is giving every new developer on longcat.ai 10 million free tokens. The model is natively multimodal, encoding images as discrete tokens in the same sequence as text, which lets it take a screenshot or a photo and then build or fix what it sees, from web pages to 3D scenes.

The engineering story is the point. Meituan says long-context fine-tuning throughput on a single domestic card rose 86 percent over a DDP baseline after the team co-designed the training stack with the chip vendor around supernode interconnects. On the inference side, a W4A4 quantization scheme with 4-bit weights and activations lifted prefill throughput 43.3 percent and decode throughput 42.77 percent against a W8A8 INT8 baseline, cut first-token wait by about 16 percent and per-token generation time by about 20 percent in online load tests. All of those figures are company-reported and have not been independently verified.

On capabilities, Meituan positions LongCat-2.5 for long-horizon agent work rather than chat. The company reports 38.5 on Terminal-Bench 4.0, up from 3.0 for LongCat-2.0, and says results across its evaluations land between Claude Opus 4.8 and Claude Opus 5, exceeding the latter on some tasks, again its own numbers. Training data for agent behavior came from real pull requests and issues in public GitHub repositories, plus failure traces from its own agents, replayed as reproducible tasks with tests attached.

The release continues a deliberate strategy. LongCat-2.0, launched June 30, was billed as the first trillion-parameter model fully trained and served on a 50,000-card domestic cluster with more than 30 trillion tokens of pretraining, and was open-sourced under MIT in July with inference adaptations from Huawei Ascend, Moore Threads and Muxi. CEO Wang Xing said on the August earnings call that Meituan will not be a token factory: the models exist to serve the core business. The LongCat stack already powers Miaoshou, an AI operations workbench for restaurant chains, and CatPaw, an internal agent platform Meituan says covers 90,000 employees and 30,000 agents.

For the industry, LongCat-2.5 is evidence that the domestic silicon question has moved from whether these chips can run a frontier model to how fast and how cheaply. A trillion-parameter model trained and quantized entirely on homegrown accelerators, with throughput gains that mostly come from software co-design, is a different argument than a paper benchmark. The open question is whether developers outside the Meituan ecosystem will adopt a model whose strongest results remain self-reported.

Comments (0)

Log in to join the discussion

Log In

No comments yet