AI Industry News: Meituan Launches LongCat-2.5; Pika Releases MiniMax H3–Based Pika H3 Flex as Open Weights

Two Sheep Co., Ltd. summarizes major developments in the AI industry based on public information such as company announcements and media reports. This article covers Meituan's native multimodal model LongCat-2.5, MiniMax's audio-video generation model MiniMax H3, and Pika H3 Flex, Pika's post-trained version of MiniMax H3. Unless otherwise noted, performance figures are as reported by the companies (as of November 3, 2026).

Key Points

• On November 1–2, Meituan announced the native multimodal model LongCat-2.5 and launched it via API. At 1.6 trillion total parameters, it matches LongCat-2.0's scale; its weights are not yet public
• Meituan says long-context training ran on Chinese domestic AI accelerators with under 80 GB of memory each, and 4-bit (W4A4) inference raised throughput by about 43% over INT8 (vendor-reported)
• In Meituan's vendor-reported benchmarks, LongCat-2.5 scores above the comparison models on BrowseComp but below them on some tests, such as DeepSWE v1.1
• On October 29, Pika announced Pika H3 Flex, a post-trained version of MiniMax's open-weight audio-video model MiniMax H3, and released adapters and inference code
• The open Pika H3 Flex package goes up to 768p and was tested on one NVIDIA H100 (80GB). The MiniMax H3 license sets conditions on territory, commercial use and more

1. Meituan's LongCat-2.5

What Was Announced

Meituan announced LongCat-2.5 in an official blog post dated November 1, and the changelog of its API platform lists the launch ("LongCat-2.5 Now Available") under November 2. The blog describes it as "a native multimodal model for long-horizon agentic tasks."

• Scale: It has 1.6 trillion total parameters, with about 48 billion activated per token. The blog says it builds on LongCat-2.0 with stronger visual understanding and agentic coding. The LongCat-2.0 model card gives the same figures, so the scale is the same as the previous model
• Uses: The changelog lists visual coding, creative design, 3D tasks and full-project builds, and says the model works with agent harnesses such as Claude Code, Codex and Hermes
• API: According to the API documentation (checked November 3), the model has a 1M-token context window and a maximum output of 128K tokens. On September 25, LongCat-2.5-Preview had already become available on the API platform, adding image understanding
• Weights: As of November 3, the weights are not public. The GitHub and Hugging Face buttons on the official blog are marked "Coming soon," the model does not appear on Meituan's Hugging Face page, and no license has been announced. The previous LongCat-2.0 launched on the API on June 30, and its weights are available on Hugging Face under the MIT license

How "Native Multimodal" Works

According to the blog, LongCat-2.5 represents visual signals such as images as multi-level discrete tokens. Image features at native resolution are turned into tokens with residual vector quantization (RVQ), the same approach as in LongCat-Next. The visual token embeddings are learned jointly with the model's backbone, and because token IDs can be precomputed offline, the blog says the vision encoder can be removed from the training graph.

Training and Inference on Chinese Domestic AI Accelerators

The blog describes training and inference on Chinese domestic AI accelerators. It does not name the vendor or product, but the page's description says the training was done entirely on domestic chips.

• Long-context training: With less than 80 GB of memory per accelerator, the training system prioritizes memory efficiency and aligns its parallelism with the "supernode" architecture, which groups multiple accelerators
• 4-bit inference (W4A4): Both the MoE expert weights and the activations are quantized to 4 bits and computed on the accelerators' INT4 units. The blog says performance was comparable to the unquantized model on all evaluated downstream tasks
• Inference performance: In online stress tests against a W8A8 (INT8) baseline, the blog reports prefill (input processing) throughput up 43.3% and decode (output generation) throughput up 42.77%, with time to first token (TTFT) down 15.91% and time per output token (TPOT) down 19.72%

Benchmarks (Vendor-Reported)

In the evaluation table, LongCat-2.5 scores above the comparison models on some tests and below them on others. The table also includes GPT-5.6-Sol, Claude Opus 4.8, GLM-5.3 and kimi-k3, among others; only some comparisons are shown here (comparison models' scores in parentheses). All scores in parentheses are marked as taken from public technical reports and similar sources; unmarked scores in the table were run by Meituan and are labeled indicative only.

• BrowseComp: 91.9 (Claude Opus 5: 90.8; GPT-5.5: 84.4)
• Terminal Bench v2.1: 88.4 (Claude Opus 5: 89.1; DeepSeek-V4.1-Flash: 90.6)
• HLE (with tools): 63.1 (Claude Opus 5: 64.7)
• DeepSWE v1.1: 64.6 (GPT-5.5: 67.0; Claude Opus 5: 68.8; DeepSeek-V4.1-Flash: 74.2)

2. MiniMax's MiniMax H3

Model Overview

MiniMax announced MiniMax H3 in an official blog post dated July 31. It said the model understands text, images, video and audio in a unified way and generates video with stereo sound, up to 15 seconds at 2K resolution (2K is provided through a workflow hosted by MiniMax). The weights were published on Hugging Face in early August (the repository's first commit was on August 4 Japan time, and the license is dated August 2).

• Architecture: According to the model card, the core is a 33-billion-parameter dense, single-stream Transformer, of which about 13 billion parameters do not need to be loaded for inference-only use. It is paired with an encoder that uses the weights of Qwen3-VL-32B to understand inputs
• Output: Clips of 4–15 seconds at 24 fps with 32 kHz stereo audio, with the shorter side set to 768 pixels by default. MiniMax says it supports dialogue in 11 languages, including Japanese
• What was not released: H3-Regenerate-2K, the module that produces 2K output, is not yet open-sourced, and H3-Context-IR, which handles preprocessing, is offered only through an API. Sparse-attention inference is also planned for a later release. As a result, the open weights alone produce 768p-class video

Main License Terms

The weights are released under the "MiniMax H3 Community License Agreement," which sets conditions on use. The main points are as follows (checked November 3).

• Territory: The license covers the world excluding the EU, the UK, the Republic of Korea and the USA. People in those territories need to contact MiniMax about a license, which MiniMax says it grants based on robust controls and safeguards for complying with those territories' laws and regulations. Japan is not excluded. Using the model or its outputs outside the covered territory is forbidden
• Commercial use: Commercial products or services with more than US$20 million (or the equivalent in other currencies) in yearly revenue need separate prior written authorization. Commercial products or services must display "MiniMax H3" prominently in their user interface
• Other terms: Outputs may not be used to improve other AI models, and military use is banned. The governing law is that of the Hong Kong SAR

If you are considering using the model, please check the full license text yourself. This article is not legal advice.

3. Pika's Pika H3 Flex

A Post-Trained Derivative of MiniMax H3

In an official blog post dated October 29, Pika announced Pika H3 Flex as "our post-trained MiniMax H3 model." Pika says it post-trained the model to generate in six steps instead of 50, and that it can generate a 15-second 768p clip with synchronized audio in as little as seven seconds of transformer (DiT) computation. A footnote explains that this figure is computation time in a controlled benchmark, not total end-to-end API latency.

What Was Released and What Stays on Pika's Service

The released items are as follows (checked November 3).

• Adapters and inference code: Two LoRA adapters (FP32, rank 128, about 5.5 GB combined) were published on Hugging Face: "FL2VA," which generates from text and endpoint frames, and "Ref2VA," which generates from multimodal references. They are distributed as Model Derivatives of MiniMax H3 under the MiniMax H3 Community License. The inference code published on GitHub is licensed under Apache-2.0
• What it takes to run: About 210 GB of MiniMax H3 base files (downloaded automatically) plus the adapters. Each request was tested on one NVIDIA H100 (80GB), but with default settings a 768p, 15-second clip exceeded H100 memory and needed CPU offload, with host memory use observed near 94 GiB (including page cache). Splitting a single request across multiple GPUs is not implemented
• Open package specs: 480p and 768p, 24 fps, 32 kHz stereo audio, 5–15 seconds per clip, and up to 12 reference files
• Pika-hosted only: The 1080p/2K/4K upsampler, separation of audio into five stems (Dialogue, Score, SFX, Foley and Ambience), prompt enhancement and the accelerated serving stack are available only through Pika's hosted services

Leaderboard and Availability

• Leaderboard: Pika said that at launch, Pika H3 Flex ranked #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard (Pika's launch snapshot). On DesignArena's video leaderboard, checked on November 3, it ranked #8 (Elo 1288), with the base MiniMax H3 at #11 (Elo 1276). That leaderboard does not show which models are open-weight, so the rank among open-weight models could not be checked
• Availability: At launch, Pika said Flex was available through the Pika API Club and would come to the Pika creation product on November 2. We could not confirm whether that rollout has taken place

4. What It Means for Data Centers and Compute

• GPU memory and storage: The open Pika H3 Flex package was tested with one NVIDIA H100 (80GB) per request, but with default settings a 768p, 15-second clip exceeded GPU memory and needed CPU offload, using about 94 GiB of host memory (including page cache); the base files alone take about 210 GB
• Higher resolutions stay hosted: MiniMax H3's 2K module has not been released, and Pika H3 Flex's output at 1080p and above is available only on Pika's service. LongCat-2.5 was also launched via API first; its weights have not yet been published
• Working within compute limits: LongCat-2.5's blog describes training on Chinese domestic accelerators with less than 80 GB of memory each and 4-bit inference, with Meituan reporting about 43% higher throughput than an INT8 setup

We believe that while work continues on running large models efficiently with limited compute, the stable operation of compute infrastructure, including GPU memory, storage, and the power and cooling behind them, will remain important as inference for models that handle video and long contexts grows.

5. About Two Sheep Co., Ltd.

Two Sheep Co., Ltd., headquartered in the Telecom Center Building in Aomi, Koto-ku, Tokyo, provides on-site work at data centers and the shipping and installation of IT equipment.

• 24/7/365 Remote Hands: at data centers in Tokyo and Osaka (Equinix TY and OS sites, AT TOKYO CC1 and CC2, NTT and others), we handle server rack-mounting, cabling, part replacement (RMA), status checks and more
• International shipping and import/export customs for IT equipment: we arrange transport and customs clearance from overseas all the way into the data center
• Migrations, connectivity and optical modules: we handle data center migrations, arrange IP transit and private lines, and sell optical modules (ModuleSheep)
• Support in Japanese, English and Chinese: we coordinate between overseas customers and Japanese data centers

Main Sources

• Meituan LongCat: LongCat-2.5 official blog (November 1, 2026, English and Chinese)
• Meituan LongCat: API platform changelog (November 2, 2026)
• Meituan LongCat: API platform documentation (checked November 3, 2026)
• Meituan LongCat: LongCat-2.0 model card (Hugging Face)
• Meituan LongCat: Public models on Hugging Face (checked November 3, 2026)
• MiniMax: MiniMax H3 announcement (July 31, 2026)
• MiniMax: MiniMax H3 model card (Hugging Face)
• MiniMax: MiniMax H3 Community License Agreement (August 2, 2026)
• Pika: Pika H3 Flex announcement (October 29, 2026)
• Pika: Pika H3 Flex model card (Hugging Face)
• Pika: Pika H3 Flex inference code (GitHub)
• Pika: Pika H3 Flex hardware and tested configurations (GitHub)
• DesignArena: Video leaderboard (checked November 3, 2026)

About This Article

• This article was compiled independently by Two Sheep from company announcements, documents such as model cards and licenses, news reports and other public information available as of November 3, 2026. Items based only on news reports are marked as such. Benchmark and performance figures are as reported by the companies and have not been verified by Two Sheep. Companies' plans and figures may change.
• This article is for information only and does not recommend any company, security or investment. Statements about licenses are not legal advice.
• Companies are named only to illustrate industry trends; their mention does not indicate any business relationship with Two Sheep. Company, product and service names are trademarks or registered trademarks of their respective owners.
• Contact: Japan@2sheep.jp, or use the contact form on this website.