LONG AI
Home › Blog › Meituan's 1.6T LongCat-2.5 Officially Op

Meituan's 1.6T LongCat-2.5 Officially Open-Sourced

Category:行业资讯 · 18 min read · 2026-11-09

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Headline: Meituan's 1.6-trillion-parameter LongCat-2.5 officially open-sourced

On Nov 9, Meituan open-sourced LongCat-2.5 under the MIT license: 1.6T parameters, 1M context, trained entirely on domestic compute, weights live on Hugging Face.

This is the largest open-weight model from a Chinese internet company, and the MIT license (free for commercial use) is the most permissive tier - a notable contrast with the "modified MIT" and "community license" approaches some peers have taken. For private deployment it's a strong candidate; the practical bar remains hardware (a 1.6T model still needs real GPU capacity to run).

Other releases (4)

Model / ProductDateTypeHighlights
NeoHorse-1.5-Flash Preview (TokenRhythm)11-9Agent-native model / limited betaTrained on Qwen3.8-Flash, 180B total / 6B active, aimed at agent task execution and software engineering. Vendor: Agents' Last Exam +9.23, Natural Language to Repository +5.92, physical-science tasks +20+. Continues RSI's "second loop": feed the previous generation back into the model pool to collect real execution feedback. Three techniques: multi-model differential diagnosis, adversarial environment evolution, swarm online learning. All self-reported, beta only, no public weights.
drvexo-4b-unified (Yujie Intelligence, open source)11-8Unified multimodal / on-deviceA "one-package" multimodal interaction model: 4B language backbone folding vision, audio input, text chat, speech generation, and candidate decision into one loadable package (6.97B total, 14.35GB). For voice comms, smart outbound calls, voice control, embodied apps; local Python + vLLM-Omni; Apache 2.0. Self-reported: MMLU-Pro 3,000-question stratified subset 77.93% (a subset, not full), plus IFEval, MATH-500, MMMU, ChartQA, JevBench. Subset scores aren't full scores; no third-party replication.
Phonon TTS (Gradium, open + non-commercial split)11-5On-device TTSTwo tiny models: 31M (English) / 40M (multilingual), ~65MB / 75MB, five languages, 100+ voices, runs on phone/browser/PC/embedded. Vendor claims German and Spanish intelligibility first on the public HF voice board. Watch the license: 31M English is Apache 2.0, the 40M multilingual is non-commercial - this week's easiest trap.
Dharma Field Extractor (Dharma AI, open source)11-4Document field extraction / 2BNot "OCR to text" but JSON to a given schema: LoRA on an OCR-pretrained base with ~43,384 labeled samples, extracting fields from invoices, contracts, IDs, tables; missing fields return null by design. Vendor: 0.834 composite over 546 docs / 3 scenarios / 1,638 tests, vs Claude Opus 5\u2019s 0.816. Self-benchmarked, and cost advantage depends on self-hosted GPU price and volume.

Industry news

The trend in one line

China's open-source flagship scale keeps climbing, now at 1.6T parameters under a permissive license.

(Data from official announcements and public reports, 2026-11-09; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa