Skip to main content
ShippedNo vapor

The AI that actually shipped

VerdictKatanemo (katanemo)

Editor's first pass (AI-assisted)

Scored by an AI research pass from the linked sources; a human editor will re-score it. How we score →

Real4.0Weights and GGUF builds are on the Hub under a Katanemo licence tagged for research use.
Useful3.5Routes by user-written domain or action descriptions rather than fixed benchmarks.
Notable3.0Powers the Arch proxy server.
Novel3.5Preference-aligned routing is the angle in its paper.
Hype gap▼ −0.5 under
Shipped

Readers

Upvotes count the readers who want this seen. Ratings are 1 to 5 scores on the four dimensions; the rating count is the number of readers who rated it. The editor’s score is shown on its own and is the headline figure. Once readers sign in, their ratings will pull a score toward the readers’ view only as more of them rate, and a reader mean from fewer than 5 ratings is too few to print. How ranking works →

Katanemo releases Arch-Router 1.5B, a router that follows your own route descriptions

For context: a Qwen2.5-1.5B-based model that maps a query to a user-defined route and returns JSON. Released under the Katanemo licence.

For context: Arch-Router is from mid-2025, outside the last twelve months, and is included as an early open router model. The card cites a paper, arXiv 2506.16655, published on 19 June 2025. The Hub repository was created on 30 May 2025.

Arch-Router is a compact 1.5B-parameter router that maps a query to a user-defined domain or action so that it can be sent to the best-suited model. It powers the Arch proxy server. It is built on Qwen2.5-1.5B. You define routes with a name and description, such as code generation or bug fixing, and pass them as JSON inside route tags along with the conversation. The model returns a JSON route name, or other if no route fits. Each route is tied to a preferred model.

The licence is the Katanemo licence, shown on the Hub with a research-use tag; the card links a LICENSE file. The Hub lists about 2B parameters in F16 where the text says 1.5B.

The card claims state-of-the-art results in matching queries with human preferences on conversational datasets, but gives no figures on the page.

A router that reads your route descriptions rather than guessing task difficulty. Read the licence before using it commercially.

Sources

Editor's first pass (AI-assisted)

Primary evidence

paperNot checked

Comments