Parallax by North ML · Invite-only

Follow a question.
See what holds up.

Meet Parallax, a research workspace that searches, reads, and checks its answers. Keep the sources in sight and follow your curiosity a little further.

A kestrel holding its wings open against the sky
Observe. Check. Answer.North ML / Kestrel

Parallax, for going deep.

A research workspace on the same models. It searches, shows which sources it kept and dropped, cites them, and saves your threads. Fiber 1 Pro checks every claim against its sources; Fiber 1 Lite is quicker.

Parallax
Which element has atomic number 26?
  1. Routing
  2. Searching
  3. Reading
  4. Verifying
  5. Writing

    Iron (Fe) has atomic number 26. [1]

    Checked against 2 sources

    An illustrative preview that runs in your browser. Try the real thing in Parallax →

    Open Parallax
    55public model repos
    36.8kall-time downloads
    Apache-2.0every Flare-1 model
    1 Mac miniserves the whole lineup

    Three models. One runtime.

    Every Flare-1 model ships with Kestrel: a router that decides whether to search or think, live retrieval, and an answer-checking pass.

    Ink 1

    ~725M parameters (565M backbone)

    Quick answers, lookups and small tasks, grounded in live sources.

    8-bit
    ~755 MB
    Weights on Hugging Face →

    Fiber 1

    ~1.63B parameters (1.41B backbone)

    More room to reason. Stronger on code and multi-step questions. Fiber 1 Lite in Parallax.

    Backbone
    Starlight Pro
    Weights on Hugging Face →

    Fiber 1 Pro

    ~1.63B parameters + verifier

    Thinks first, then checks every claim and number against its sources and rewrites what doesn't hold up.

    Where
    Parallax only
    Training
    Online GRPO
    Weights on Hugging Face →

    It hovers before it strikes.

    Our runtime is named after the kestrel, the falcon that holds still in the wind, looks, and only then commits. Every grounded answer does the same.

    1. RouteA router head decides: search, think, or just answer.
    2. SearchWikipedia and the live web, reranked for relevance.
    3. VerifyFiber 1 Pro checks claims and numbers against the sources.
    4. WriteCited answer, streamed stage by stage.

    Drop-in OpenAI compatible.

    Point the OpenAI SDK at North ML and keep your code. Streaming, usage reporting, and the stages the model went through, all in the same response.

    • /v1/chat/completions with stream: true
    • north-auto routes each request to Ink or Fiber
    • Keys, usage and logs in a real dashboard
    Read the docs
    import os
    from openai import OpenAI
    
    client = OpenAI(
        base_url="https://www.north-ml.space/v1",
        api_key=os.environ["NORTH_API_KEY"],
    )
    
    stream = client.chat.completions.create(
        model="north-auto",
        messages=[{"role": "user", "content": "What do kestrels eat?"}],
        stream=True,
    )
    for chunk in stream:
        print(chunk.choices[0].delta.content or "", end="")

    Published, failures included.

    If we publish a number, you can download the thing that produced it and check it yourself.

    Run it on your own desk.