LongCat-2.5 open weights: MIT license, $0.30/$1.20 API

Another frontier-class open model just landed, and the license is MIT. Meituan released the LongCat-2.5 open weights on November 9, one week after it introduced the model, VentureBeat reports. It's a native multimodal mixture-of-experts model built for long agent tasks. The hosted API is listed at $0.30 per million input tokens and $1.20 per million output, at a limited-time discount. Here's what's real and what to check.

What actually happened

• The model. Per the Hugging Face model card, LongCat-2.5 has 1.6 trillion total parameters with about 48 billion active per token. It reads images and video along with text. VentureBeat notes that Hugging Face's metadata says 1.8T, so size any deployment from the actual files.
• The license. The weights are MIT. The model card adds that this doesn't grant rights to Meituan's trademarks or patents.
• The price. Meituan's pricing page lists $0.30 per million uncached input tokens, $0.006 cached and $1.20 output. It's marked "Discounted Price (limited-time)," with no end date, and "Prices are subject to change." The API speaks Chat Completions, Responses and Anthropic Messages formats.
• The benchmarks. Meituan's card puts LongCat-2.5 at 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, close to the Claude and GPT models it compares against. The card says its own runs are "measured in-house and are indicative only." VentureBeat also points out that the comparisons use older rival models, not the current ones.
• Self-hosting. This isn't a one-GPU model. VentureBeat's rough math puts the 16-bit weights alone at about 3.2 TB. Meituan also ships a 4-bit W4A4 version. Hugging Face lists no third-party inference provider for it yet.

Why it matters for your business

Cheap, capable open models are now a monthly event. That's good for you, because every one of them is a price ceiling on the closed APIs. But three things should stop you from switching on a press release:

1. The price is a promo. Model your costs at a number higher than $0.30/$1.20. Discounts with no end date end anyway.
2. The benchmarks are the vendor's. Run your own tasks: your support tickets, your product data, your tool calls. Thirty real examples beat any leaderboard.
3. Know where your data goes. The cheap price is on Meituan's own API. If customer data can't leave your jurisdiction or your approved vendors, wait for a provider you've already vetted to host the MIT weights, or host them yourself.

The MIT license is the durable part. Weights you can download today can't get repriced on you tomorrow.

Sources: VentureBeat, Hugging Face model card, LongCat pricing.