LongCat

LongCat-2.0 and Qwen3.8-Max: Two Chinese Open-Weight MoE Flagships Compared

A source-based comparison of LongCat-2.0 (1.6T MoE, MIT License, 1M context, domestic-ASIC training) and Qwen3.8-Max (2.4T MoE, 95B active, custom license with revenue-share clause, open weights released August 12). Covers architecture, licensing, benchmarks, pricing, and deployment — the first LongCat vs Qwen comparison on this site.

Independent third-party resource. Not affiliated with or endorsed by LongCat, Meituan, DeepSeek, or any other publisher discussed on this page.

Published: 2026-08-20 · Author: LongCat Community Hub editorial team

This is an independent, source-based comparison of LongCat-2.0 (Meituan) and Qwen3.8-Max (Alibaba) — the first LongCat vs Qwen comparison on this site. It compares documented architecture, licensing, benchmarks, pricing, and deployment options. It does not report performance as a verdict — for publisher-reported scores, see the individual model pages and benchmark index linked below. Every claim on this page is sourced from the linked publisher documentation or attributed third-party coverage.

DimensionLongCat-2.0Qwen3.8-Max
DeveloperMeituan (LongCat team)Alibaba (Qwen team)
Release dateJune 2026August 3, 2026 (API); open weights August 12
Total parameters1.6T (MoE)2.4T (MoE)
Active parameters~48B per token (33B–56B dynamic)~95B per token
Context window1M tokens (native LSA)1M tokens (API); open weights text-only without native 1M context
ArchitectureScMoE + LSA + N-gram Embedding + MOPDSparse MoE + hybrid attention (documented)
Training hardwareEnd-to-end domestic AI ASICs (~50K cluster)Not published (NVIDIA reported for prior models)
LicenseMIT (open weights, no conditions)Custom Qwen3.8-Max License (revenue-share clause above $50M; attribution above 100M MAU)
Open weights scopeFull model; OpenAI + Anthropic compatible endpointsText-only, thinking-only variant; vision/1M context remain API-only
Benchmarks (independent)Publisher-reported SWE-bench Pro 59.5, Terminal-Bench 2.1 70.8Independent: Terminal-Bench 2.1 86.6%, AA Intelligence Index 55.4
API pricing (per 1M tokens)$0.30 input / $1.20 output (OpenRouter)$2.00 input / $6.00 output
DeploymentSelf-hosted (SGLang, vLLM, NPU path); consumer-friendlySelf-hosted requires data-center hardware (BF16 ~4.9TB); reference: 72-GPU GB300 rack

Two Chinese Open-Weight Flagships, Two License Philosophies

Qwen3.8-Max is Alibaba's largest model ever shipped: 2.4T total parameters with ~95B active, announced August 3, 2026 with a 1M-token context and multimodal (text + vision) API, and released as open weights on August 12 under a custom Qwen3.8-Max License. LongCat-2.0 is Meituan's 1.6T MoE flagship, released June 2026 under the MIT License with full end-to-end training on domestic AI ASICs.

The most consequential difference is licensing. LongCat-2.0's MIT license is unconditional: the weights can be used, modified, and redistributed freely. Qwen3.8-Max's open weights carry a custom license that requires companies running a Model-as-a- Service or AI-assistant business above $50M in annual revenue to obtain a separate license from Qwen, and requires prominent model attribution above 100M monthly active users. For most developers the open weights are usable, but for large inference providers the two models have materially different commercial terms.

The Open-Weight Release: Not Identical to the API

A second difference that matters for deployment planning: the Qwen3.8-Max open weights are not identical to the hosted API model. The released Qwen3.8-2.4T-A95B checkpoint is text-only and thinking-mode-only — it does not include the API version's vision input or its native 1M-token context. LongCat-2.0's open weights, by contrast, match the API model (both are 1M context, and the model exposes OpenAI- and Anthropic-compatible endpoints).

Alibaba also released the Qwen3.8-27B companion on August 13 under Apache 2.0 — a dense 27B multimodal model with 262K native context, aimed at local hardware. That is a genuinely permissive open release, but it is a different, much smaller model than the 2.4T flagship. Teams choosing "open weights" therefore face a decision LongCat-2.0 users do not: the flagship's weights carry conditions and omit API features, while the fully permissive release is a different size class.

Benchmarks: Different Suites, Don't Cross-Compare Blindly

Qwen3.8-Max has accumulated independent third-party measurements since its August 3 launch: Terminal-Bench 2.1 at 86.6%, an Artificial Analysis Intelligence Index of 55.4, and an Agentic Index position ahead of Claude Opus 5 and GPT-5.6 per Artificial Analysis. LongCat-2.0's published scores are publisher-reported (SWE-bench Pro 59.5, Terminal-Bench 2.1 70.8) and have not been independently verified on the same suites.

These numbers come from different evaluation suites and different measurement regimes (independent vs publisher-reported), so direct comparison is misleading. What the evidence supports: Qwen3.8-Max is a credible frontier-class model with strong independently measured agentic results, while LongCat-2.0's positioning relies on publisher claims plus its distinctive domestic-ASIC training story. Evaluate on the benchmark closest to your production task rather than comparing headline scores.

Pricing and Deployment Realities

On the API, LongCat-2.0 is priced at $0.30/$1.20 per M tokens on OpenRouter against Qwen3.8-Max's $2.00/$6.00 — a roughly 6.7x input and 5x output difference. Self-hosting changes the picture differently for each model: LongCat-2.0 documents consumer- friendly deployment paths (SGLang, vLLM, plus an NPU path for domestic accelerators), while Qwen3.8-Max's 2.4T weights are a data-center proposition — the BF16 checkpoint is on the order of 4.9TB, and Alibaba's reference deployment is a 72-GPU GB300 rack, with aggressively quantized GGUF variants still in the hundreds of GB.

For teams that need frontier-scale agentic coding with independently verified results and can operate data-center hardware, Qwen3.8-Max is the stronger documented option. For teams that need unconditional open licensing, native 1M context in the open weights, domestic-hardware training alignment, or dramatically lower API cost, LongCat-2.0 is positioned differently. The two models serve overlapping but not identical decisions.

This comparison is based on publicly available documentation accessed on 2026-08-20. LongCat-2.0 specifications are from the publisher blog and GitHub repository. Qwen3.8-Max specifications are from Alibaba's announcement, the HuggingFace release, and attributed independent analyses of the license and benchmarks.

This is a feature-level comparison, not a performance evaluation. This site has not independently tested either model. LongCat-2.0 benchmark figures are publisher-reported; Qwen3.8-Max independent figures are attributed to Artificial Analysis and other third-party platforms. The two models' scores come from different suites and are not directly comparable.

Pricing, licensing, and availability are current as of the access date and may change. Verify the full Qwen3.8-Max License terms and current rates on each vendor's official documentation before deployment.

Related pages

Related comparisons in this series

  • LongCat-2.0 and Kimi K3: Open-Weight MoE Flagships Compared

    A source-based comparison of LongCat-2.0 (1.6T MoE, MIT License, domestic-ASIC training, $0.30/M input) and Kimi K3 (2.8T MoE, modified MIT License, native vision, $3.00/M input) — the two largest open-weight Chinese models. Covers architecture, licensing, pricing, modalities, and deployment.

  • LongCat-2.0 and DeepSeek V4-Flash: Open-Weight MoE Models Compared

    A source-based comparison of LongCat-2.0 (1.6T MoE, ~48B active, MIT License, $0.30/M input) and DeepSeek V4-Flash (284B MoE, 13B active, MIT License, $0.14/M input). Covers architecture, pricing, cache economics, deployment, and licensing.

  • LongCat-2.0 and GLM-5.2: Open-Weight MoE Flagships Compared

    A source-based comparison of LongCat-2.0 (1.6T MoE, ~48B active, MIT License, $0.30/M input) and GLM-5.2 (744B MoE, ~40B active, MIT License, $1.40/M input). Covers architecture, pricing, licensing, coding benchmarks, and deployment.

Sources

  • LongCat-2.0 Publisher Blog Post

    Publisher documentationAccessed 2026-08-20

    Publisher announcement detailing LongCat-2.0 architecture (ScMoE, LSA, N-gram Embedding, MOPD), 1.6T parameters, domestic-ASIC training, and MIT License.

  • LongCat-2.0 GitHub Repository

    Primary sourceAccessed 2026-08-20

    Model code, weights, MIT License terms, and deployment instructions.

  • Qwen3.8-Max official blog (August 3)

    Publisher documentationPublished 2026-08-03Accessed 2026-08-20

    Official announcement: 2.4T total / 95B active MoE, 1M context, autonomous coding case studies (16-day run, 125-hour paper reproduction), benchmark claims, API availability.

  • Qwen3.8-2.4T-A95B on HuggingFace

    Primary sourcePublished 2026-08-12Accessed 2026-08-20

    Open-weight release August 12: BF16 + FP8 variants under the custom Qwen3.8-Max License. Text-only, thinking-only; no vision or native 1M context in open weights.

  • Qwen3.8-Max License analysis (SQ Magazine)

    Third-partyPublished 2026-08-16Accessed 2026-08-20

    License analysis: MaaS/AI-assistant businesses above $50M annual revenue must obtain a separate license; products above 100M MAU or $20M monthly revenue must display the model name prominently.

  • Qwen3.8-Max independent benchmarks (aimadetools, August 2026)

    Third-partyAccessed 2026-08-20

    Independently measured: Terminal-Bench 2.1 86.6%, Artificial Analysis Intelligence Index 55.4, Agentic Index ahead of Opus 5 and GPT-5.6; API pricing $2/$6 per M tokens.

Independent third-party disclosure

This page is published by an independent third-party site. It is not affiliated with, endorsed by, sponsored by, or operated by Meituan, LongCat, or any of their affiliates. The content summarizes publicly-available primary documentation and does not represent the views of any referenced organization.

Read the disclaimerAbout this site

Last reviewed: 2026-08-20