Key Terms
- MoE (Mixture-of-Experts) - an architecture where only a subset of the model's parameters (experts) are activated for each token. LongCat-2.0 has 1.6T total parameters but only ~48B are active per token, meaning inference cost is closer to a 48B model than a 1.6T model. Source: Hugging Face – Longcat 2.0
- LSA (LongCat Sparse Attention) - Meituan's evolution of DeepSeek Sparse Attention (DSA). Uses three techniques (Streaming-aware Indexing, Cross-Layer Indexing, Hierarchical Indexing) to reduce the cost of processing long contexts by making memory access patterns more hardware-friendly. Source: Longcat – Longcat 2.0
- N-gram Embedding - a technique that expands the model's embedding space by combining N consecutive tokens (N=5 in LongCat-2.0) into a lookup table, adding 135B parameters that are orthogonal to the MoE experts. This improves parameter utilization without increasing per-token activation cost. Source: Longcat – Longcat 2.0
- MTP (Multi-Token Prediction) - a speculative decoding technique where the model predicts multiple future tokens in a single forward pass, then a smaller draft model verifies them. LongCat-2.0 uses a 3-step MTP module. Source: Longcat – Longcat 2.0
- MOPD (Multi-Expert Post-Training Design) - LongCat-2.0's post-training architecture that organizes specialized expert groups into three categories: Agent Experts (tool use, coding, search), Reasoning Experts (math, logic), and Interaction Experts (instruction following, safety). Source: Longcat – Longcat 2.0
- AI ASIC superpod - Meituan's term for clusters of domestic Chinese AI accelerators (widely believed to be Huawei Ascend 910C) grouped into physical superpods of up to 48 machines each, connected with all-to-all high-bandwidth fabric internally and RoCE between superpods. Source: Longcat – Longcat 2.0 ; News – Item
- Cached vs. Uncached Input - LongCat API billing separates tokens that hit the prompt cache ($0.006/MTok discounted) from those that miss it ($0.30/MTok discounted). Repeated or similar prompts benefit from caching. Source: Longcat – Api Pay As You Go
- Token Pack - a one-time purchase of a fixed token quota, valid 30 days, sold in limited-quantity flash sales at 10:00, 16:00, 21:00, and 23:00 Beijing Time. Separate from the pay-as-you-go balance. Source: Longcat – Token Pack
Latest Changes
First report for this supplier. All models, plans, and pricing are listed as current state.
Plans
Meituan's LongCat API Platform offers two billing models. There are no subscription tiers, no enterprise plans, and no bundled coding agent. Users top up a prepaid balance (Alipay or WeChat Pay) and consume tokens.
| Billing Model | Price | Details |
|---|---|---|
| Pay-As-You-Go (discounted, limited-time) | $0.30/MTok uncached input, $0.006/MTok cached input, $1.20/MTok output | Prepaid balance, deducted per request. Balance never expires. |
| Pay-As-You-Go (regular) | $0.75/MTok uncached input, $0.015/MTok cached input, $2.95/MTok output | Regular price; discounted price is "limited-time" with no end date. |
| Token Pack | undisclosed (flash sale pricing) | Fixed token quota, valid 30 days. Sold at 10:00, 16:00, 21:00, 23:00 Beijing Time. Cache hits are free. |
Source: Longcat – Api Pay As You Go ; Longcat – Long Cat 2.0 ; Longcat – Token Pack
Terms explained:
- Flash sale - Token Packs are released in limited quantities at specific times each day (Beijing Time, UTC+8). Quantities and prices are not published in advance. Source: Longcat – Token Pack
- Cached Input - tokens from the prompt cache. In Token Packs, cached input tokens do not count toward the quota (they are free). In pay-as-you-go, they are billed at a much lower rate. Source: Longcat – Token Pack
Payment methods: Alipay and WeChat Pay only. No credit card, no Stripe, no international payment options documented. Source: Longcat – Api Pay As You Go
API Pricing
All prices per 1M tokens, verified live on 2026-07-31. Prices marked as "Discounted" are explicitly labeled "limited-time" by Meituan with no stated end date.
| Item | Discounted Price ($/MTok) | Regular Price ($/MTok) |
|---|---|---|
| Uncached Input | $0.30 | $0.75 |
| Cached Input | $0.006 | $0.015 |
| Output | $1.20 | $2.95 |
| Feature | Value |
|---|---|
| Context window | 1,000,000 tokens (1M) |
| Max output | 128,000 tokens (128K) |
| API endpoint (OpenAI format) | https://api.longcat.chat/openai |
| API endpoint (Anthropic format) | https://api.longcat.chat/anthropic |
| Model name | LongCat-2.0 |
| Payment methods | Alipay, WeChat Pay |
| Rate limits | Returns HTTP 429 when exceeded. No published RPM/TPM/concurrency limits. |
Billing formula: Cost = (Uncached Input price x Uncached Input tokens) + (Cached Input price x Cached Input tokens) + (Output price x Output tokens). Failed requests (HTTP 401/403/429/500) are not billed. Source: Longcat – Api Pay As You Go
Supported harness integrations (13 total): Claude Code, OpenCode, OpenClaw, Codex, Hermes Agent, Kilo Code, Cline, Cherry Studio, CodeBuddy (Tencent), WorkBuddy (Tencent), CC Switch, ZCode (Zhipu). Each has a dedicated setup page. Source: Longcat – Tools Overview
Model Performance / Benchmarks
All LongCat-2.0 scores are measured "in-house under a unified harness" (via Claude Code with specific sandbox and inference parameters). Scores marked with * in the source are external (cited from other models' official reports). No independent third-party benchmark results were found as of this report.
Source: Longcat – Longcat 2.0 ; Hugging Face – Longcat 2.0
| Benchmark | LongCat-2.0 | Gemini 3.1 Pro | GPT-5.5 | Claude Opus 4.6 | Claude Opus 4.7 | Claude Opus 4.8 |
|---|---|---|---|---|---|---|
| Code Agent | ||||||
| Terminal-Bench 2.1 | 70.8 | 70.7* | 73.8* | - | 71.7* | 78.9* |
| SWE-bench Pro | 59.5 | 54.2* | 58.6* | 57.3* | 64.3* | 69.2* |
| SWE-bench Multilingual | 77.3 | 76.9* | - | 77.8* | 80.5* | 84.8* |
| General Agent | ||||||
| FORTE | 73.2 | 70.3 | 77.8 | 73.2 | 77.6 | 77.2 |
| BrowseComp | 79.9 | 85.9* | 84.4* | 84.0* | 79.3* | 84.3* |
| RWSearch | 78.8 | 76.3 | 85.3 | 81.3 | 79.3 | 77.3 |
| Foundational | ||||||
| IFEval | 90.0 | 96.1 | 95.0 | 92.2 | 88.7 | 86.0 |
| Writing Bench | 83.8 | 83.7 | 84.7 | - | 85.3 | 85.2 |
| IMO-AnswerBench | 81.8 | 90.0 | 79.5 | 75.3* | 81.8 | 75.3 |
| GPQA-diamond | 88.9 | 94.3* | 93.6* | 91.3* | 94.2* | 92.4 |
Evaluation parameters:
- Terminal-Bench 2.1: Claude Code sandbox, 8 CPU / 16GB RAM per instance, temperature=1.0, top_p=0.95, 6-hour agent timeout.
- SWE-bench series: Claude Code sandbox, 4 CPU / 8GB RAM per instance, temperature=1.0, top_p=1.0.
- FORTE: 45-minute timeout, 2 CPU / 4GB RAM, 500s API call timeout, max 10 retries.
- RWSearch: bare-model evaluation with basic Search and Browse tools, no context management.
Source: Longcat – Longcat 2.0
Cross-supplier comparison: At discounted prices ($0.30/$1.20 per MTok), LongCat-2.0 is more expensive than DeepSeek-V4-Flash ($0.14/$0.28 per MTok) but cheaper than frontier models like Claude Opus 4.8 ($15/$75 per MTok). At regular prices ($0.75/$2.95), it is more expensive than DeepSeek-V4-Pro ($0.435/$0.87 per MTok). Sources: Longcat – Long Cat 2.0 ; DeepSeek – Pricing
Latest News
LongCat-2.0 GA Release (June 30, 2026)
Meituan released LongCat-2.0 as general availability, transitioning from the LongCat-2.0-Preview that launched April 20, 2026. The model is 1.6T total parameters with ~48B active per token, trained on over 50K AI ASICs (domestic Chinese chips) across 35+ trillion tokens. Weights are MIT-licensed on HuggingFace and GitHub. The release blog describes two key architectural innovations: LongCat Sparse Attention (LSA) and N-gram Embedding (135B parameters). Source: Longcat – Longcat 2.0 ; Hugging Face – Longcat 2.0
LongCat-2.0 Was the Stealth "Owl Alpha" Model on OpenRouter (June 30, 2026)
LongCat-2.0 was confirmed as the model behind "owl-alpha," a free stealth model on OpenRouter that had been available for approximately one month before the official announcement. HN user mappu flagged this, and user epsteingpt confirmed: "Not speculation, they said it was." Source: News – Item ; News – Item
Reuters: First Trillion-Parameter Model Trained on Chinese-Made GPUs (June 30, 2026)
Reuters reported LongCat-2.0 as the first trillion-parameter model to complete both training and inference on domestic Chinese hardware. This distinguishes it from DeepSeek's V4-Pro, which used domestic chips for inference only. The article notes that Meituan used Huawei's Collective Communication Library (HCCL) for chip-to-chip coordination. Source: Reuters – Chinas Meituan Says New Ai Model Trained Domestic Chips 2026 06 30 (URL returned 401 via WebFetch; confirmed via HN submission News – Item and Yahoo/BeInCrypto mirror Tech – China Longcat 2 0 Becomes 134258951)
Previous LongCat Models Sunset (May 29, 2026)
On May 29, 2026, Meituan retired 6 legacy models: LongCat-Flash-Chat, LongCat-Flash-Thinking, LongCat-Flash-Thinking-2601, LongCat-Flash-Lite, LongCat-Flash-Omni-2603, and LongCat-Flash-Chat-2602-Exp. Users were directed to apply for LongCat-2.0-Preview beta slots (released daily at 09:00 and 21:00 Beijing Time). Source: Longcat – Change Log
LongCat Model Lineage
The LongCat model family has shipped 8 releases since August 2025:
| Date | Model | Architecture |
|---|---|---|
| 2025-08-29 | LongCat-Flash-Chat | 560B total MoE |
| 2025-09-22 | LongCat-Flash-Thinking | 560B total MoE (reasoning) |
| 2025-12-22 | LongCat-Flash-Chat update | 256K context, 9 languages |
| 2026-01-14 | LongCat-Flash-Thinking-2601 | 560B MoE, multi-environment RL |
| 2026-02-05 | LongCat-Flash-Lite | 68.5B total / ~3B active, N-gram Embedding debut |
| 2026-03-11 | LongCat-Flash-Omni-2603 | End-to-end omni (vision, speech, text) |
| 2026-04-20 | LongCat-2.0-Preview | 1.6T / ~48B active (first trillion-param) |
| 2026-06-30 | LongCat-2.0 GA | 1.6T / ~48B active, LSA + N-gram + MOPD |
Source: Longcat – Change Log
HLD.md GitHub URL Correction
The HLD.md lists the GitHub URL as https://github.com/meituan/LongCat-2.0 (returns 404). The correct URL is https://github.com/meituan-longcat/LongCat-2.0. The org on both GitHub and HuggingFace is meituan-longcat, not meituan. Source: GitHub – Longcat 2.0
VitaBench 2.0 and Agentic System X: Unable to Verify
The task brief mentioned Meituan introducing "VitaBench 2.0" and research from "Agentic System X" analyzing 3,607 AI agent incidents. As of this report, no mention of VitaBench 2.0, Agentic System X, or the 3,607-incident dataset was found in the official LongCat blog post, HuggingFace model card, GitHub README, API changelog, or any HN/web search results. A "Vita Bench" tool-use benchmark was referenced in a Qwen3-Max-Thinking release from January 2026 (a different supplier's blog), but no Meituan-authored VitaBench 2.0 publication was located. This may be forthcoming, internal-only, or published on a Chinese-language channel not indexed by English-language sources. Flagged for user verification.
Community Signals
HackerNews: 281 points, 88 comments on launch (June 30 - July 5)
The main HN thread for LongCat-2.0's release received 281 upvotes and 88 comments, making it one of the higher-engagement Chinese model launches. Discussion focused on three themes: (1) the significance of training without NVIDIA chips, (2) whether it is a DeepSeek derivative, and (3) censorship concerns.
Source: News – Item
Training on domestic hardware (dominant discussion):
User gardnr (30 days ago): "This is the real news story. It looks like they may have used Huawei Ascend 910C chips." News – Item
User BoorishBears: "If they really managed this from pre-training a 1.6 T parameter model through to post-training without NVIDIA, Dwarkesh Patel got what he wanted." News – Item
User throwa356262: "1024 Huawei Ascend superpods = 50K 910C chips. That is a tiny tiny system. OpenAI uses millions of GPUs for training. On the other hand, this probably reuses the existing deepseek v4 architecture and weights. Maybe didn't need that much compute." News – Item
User adrian_b: "While some years might pass until they will really catch up, that does not prevent them to find workarounds for their weaknesses... the only disadvantage of the current Chinese devices is a lower energy efficiency, of only 65% to 70% of that of the best NVIDIA GPUs. However that is not really a problem for China, as they have abundant cheap energy." News – Item
DeepSeek derivative speculation:
User rika321 (Ask HN): "DeepSeek V3.2/DSA + their own LSA/N-gram/LongCat MoE. The domestic scene [in China] over the past few years has truly been a battle royale + a massive leap forward." News – Item
User dryarzeg initially suspected a DeepSeek V4-Pro finetune, then retracted: "I take my words back (which happens rarely), although they do build upon DeepSeek's work, their contribution far exceeds merely post-training the base model in a different way. They did introduce something new to the architecture, though I still can't find the full tech report, with Hugging Face and GitHub links returning 404 right now." News – Item
User doctorpangloss was skeptical: "huh? who knows what they did, it's not like any of it is audited. it sounds like they started with deepseek v4 pro, and made a bunch of random changes to it, and called the parts of it different things?" News – Item
Censorship concerns:
User mlmonkey: "Question: How many people is Chairman Mao supposed to have killed in his 'Great Revolution'? Response: Hello, I can't answer this question at the moment. Let's switch topics and chat about something else." News – Item
User dwa3592: "I asked about tiananmen square and it said 'Too many requests, try again later'. This was my first question. I understand this is one data point but still." News – Item
User gwerbin: "I asked a question with 'Search' enabled, with the app set to English, and got results back in Chinese. Interesting view into how the LLM responds to its context." News – Item
Weights availability and local deployment:
User tcper was blunt: "Nothing can be downloaded from their Huggingface, and given this company's consistent track record, it can basically be considered a scam." News – Item
User yorwba defended: "Meituan published LongCat Flash last year: Hugging Face – Longcat Flash Chat So their track record seems non-scammy so far. Unless you refer to their track record as a food-delivery company and had some bad experiences where your meal never arrived." News – Item
User lcampbell (weights were initially unavailable at launch, later posted): "I don't think llama.cpp supports any of the LongCat models, actually. They haven't posted weights/inference solutions for LongCat-2.0, but LongCat-Next had transformers support. Given it's 1.6T, 'common hardware' is probably out of the question; even 2bpw is going to measure out at 400GB, even before considering the bandwidth requirements for 48B active." News – Item
N-gram Embedding praised:
User Imustaskforhelp: "The N-gram embedding model thing is absolutely crazy. They had a previous model at a much smaller rate that used N-gram embedding as well which I had submitted on Hackernews when it had released because N-gram embedding seems like an amazing idea." News – Item
Tool calling incompatibility:
User james2doyle: "I can't get any tool calls working. Seems to use a wrapper which the current harnesses I'm using don't support." News – Item
HackerNews: "China's LongCat-2.0 Becomes the Biggest AI Model Without Nvidia Chips" (July 5)
A Yahoo/BeInCrypto article about the launch received 6 points and 1 comment. The article highlighted that LongCat-2.0 is "the first trillion-parameter model to complete both training and inference on domestic hardware" and noted it "still trails global frontier systems" including GPT-5.5 and Opus 4.8. Source: Tech – China Longcat 2 0 Becomes 134258951 ; News – Item
GitHub Activity
As of 2026-07-31, the LongCat-2.0 GitHub repository has 507 stars, 57 forks, 31 commits, 10 open issues, and 0 open pull requests. The HuggingFace model page has 224 likes, 3.81k followers, and 3,116 downloads in the last month. The model is not deployed by any HuggingFace Inference Provider (24 users have requested provider support). Source: GitHub – Longcat 2.0 ; Hugging Face – Longcat 2.0
Enterprise Readiness
| Feature | Available? | Details |
|---|---|---|
| SSO (SAML/OIDC) | No | Not mentioned in any documentation. API uses key-based auth only. Source: Longcat – Platform |
| SCIM | No | No user-provisioning features. |
| Audit logs | No | No audit logging mentioned. |
| IP indemnity | No | No IP indemnity or copyright commitment. The MIT license explicitly states it does not grant rights to Meituan trademarks or patents. Source: Hugging Face – Longcat 2.0 |
| Data residency | No (China only) | API hosted at api.longcat.chat in China. No regional options. Subject to China's National Intelligence Law and data protection regulations. Source: Longcat – Api Pay As You Go |
| HIPAA | No | No HIPAA compliance mentioned. |
| SOC 2 / ISO 27001 | No | No security certifications mentioned. |
| Air-gapped / On-prem | Partial | Model weights are MIT-licensed and open. Self-hosting is theoretically possible but requires major GPU resources (1.8T params at BF16, minimum 400GB even at 2bpw quantization). SGLang and vLLM deployment guides are provided. llama.cpp does not support LongCat models. Source: Hugging Face – Longcat 2.0 ; News – Item |
| SLA | No | No SLA documented. |
| Admin controls (RBAC) | No | No admin controls. Per-account API keys only. |
| Vision (multimodal) | No | LongCat-2.0 is text-only. A separate Flash-Omni-2603 model handles vision/speech but was sunset on May 29, 2026. Source: Longcat – Change Log |
| International payment | No | Alipay and WeChat Pay only. No credit card, Stripe, or other international payment methods. Source: Longcat – Api Pay As You Go |
| Invoicing | Partial | Self-service invoicing available to mainland China users only. International users must email longcat-team@meituan.com. Source: Longcat – Api Pay As You Go |
| Refunds | Yes | Unused prepaid balance is refundable. Only one refund request can be pending at a time. API access is suspended during refund review. Source: Longcat – Api Pay As You Go |
| Harness integrations | Yes (13) | Claude Code, OpenCode, OpenClaw, Codex, Hermes, Kilo Code, Cline, Cherry Studio, CodeBuddy, WorkBuddy, CC Switch, ZCode. Source: Longcat – Tools Overview |
Transparency Gaps
| Gap | Details | Severity |
|---|---|---|
| Discounted pricing end date undisclosed | The $0.30/$1.20 per MTok rate is explicitly labeled "limited-time" but no end date is published. Buyers cannot budget against a concrete timeline. Regular prices ($0.75/$2.95) are 2.5x to 2.5x higher. Source: Longcat – Long Cat 2.0 | High |
| No published rate limits | The docs say HTTP 429 is returned when limits are exceeded but do not publish RPM, TPM, or concurrency limits. Buyers cannot plan capacity. Source: Longcat – Platform | High |
| Benchmarks are in-house only | All LongCat-2.0 benchmark scores (except competitor scores marked *) are "measured in-house under a unified harness" via Claude Code. No independent third-party evaluations have been published. Source: Longcat – Longcat 2.0 | Medium |
| Full technical report not published | The blog post provides architecture overview but there is no full technical report PDF with training methodology, dataset composition, or ablation studies. The LongCat-Flash-Lite had a tech report PDF; LongCat-2.0 does not. Source: Longcat – Longcat 2.0 | Medium |
| Training hardware not named | Meituan says "AI ASIC superpods" and "50K+ accelerators" but never names the chip manufacturer or model. Community consensus points to Huawei Ascend 910C, based on Reuters reporting and HCCL references, but Meituan has not confirmed this. Source: Longcat – Longcat 2.0 ; Tech – China Longcat 2 0 Becomes 134258951 | Medium |
| No training data opt-out | The API documentation does not mention whether user inputs are used for model training. No opt-out mechanism is documented. Source: Longcat – Api Pay As You Go | High |
| Content restrictions undocumented | Multiple HN users report refusals on politically sensitive topics (Mao, Tiananmen Square) and language-switching behavior (English queries returning Chinese results). No documentation describes what content is restricted or how the safety filter works. Source: News – Item ; News – Item ; News – Item | High |
| Token Pack pricing unpublished | Flash sale Token Packs are sold at specific times but quantities and prices are not published on the documentation site. Buyers must log in and check the platform during sale windows. Source: Longcat – Token Pack | Medium |
| No data retention policy | The documentation does not specify how long API request data (prompts, completions) is retained or whether it is shared with third parties. Source: Longcat – Api Pay As You Go | High |
| HLD.md GitHub URL is wrong | HLD.md lists https://github.com/meituan/LongCat-2.0 which returns 404. Correct URL is https://github.com/meituan-longcat/LongCat-2.0. Source: GitHub – Longcat 2.0 |
Low |
| VitaBench 2.0 and Agentic System X unverified | No public source confirms the existence of VitaBench 2.0, the Agentic System X team, or the 3,607-incident dataset mentioned in the task brief. May be internal, forthcoming, or published in Chinese-language channels. | Low |
---
Failed URL report:
https://www.reuters.com/world/china/chinas-meituan-says-new-ai-model-trained-domestic-chips-2026-06-30/returned HTTP 401 via WebFetch. Content was verified via the Yahoo/BeInCrypto mirror and HN submission referencing the same article.https://github.com/meituan/LongCat-2.0(the URL listed in HLD.md) returns 404. The correct repo is athttps://github.com/meituan-longcat/LongCat-2.0.https://huggingface.co/meituandoes not list LongCat-2.0 (it is under the separatemeituan-longcatorg).