COLDTENSOR

Advanced AI shouldn't need a data center.

Open research on making strong models run anywhere, with every tradeoff shown.

See releases Hugging Face X
Screen printed illustration of a faceless figure in a coat, its head replaced by a stack of glowing cobalt slab layers, textured print on cream paper.

What we do

We shrink strong models and publish what happened.

We take open models and make them smaller and faster to run, then release the result with the full record: what worked, what broke, and what it cost.

01

Every tradeoff shown

Size, speed, and accuracy are published together, not cherry picked.

02

Numbers published even when bad

A failed run gets a number too, not a quiet rewrite.

03

Open weights

Weights are public to download, not a demo behind a form.

04

What broke goes in the record

Failed attempts are listed next to the one that worked.

How we work

Four methods. We mix and match.

ColdTensor Firn series mark: a tall stack of cobalt slabs shrinking to a short stack, representing distillation.

Firn

Distillation

Train a small model to match a larger one.

ColdTensor Rime series mark: a stack of cobalt slabs with frost like marks at the edges, representing quantization.

Rime

Quantization

Store numbers with less precision.

ColdTensor Calve series mark: a stack of cobalt slabs with one layer shown as a dashed outline, representing pruning.

Calve

Pruning

Remove the parts a model rarely uses.

ColdTensor Floe series mark: a stack of cobalt slab layers on cream paper, flat graphic style.

Floe

Combined methods

Stack two or more of the above.

Releases

The log. Newest first.

  • FLOE-01

    Floe series
    FLOE-01 spec card: base Moonshot AI Moonlight 16B A3B Instruct, method prune 25 percent plus mixed 3 to 8 bit plus quantization aware training, size 6.71 GB from a 31.93 GB base, score 427 of 500 on GSM8K holdout versus 432 for F16, license MIT.

    Base: Moonlight-16B-A3B-Instruct (Moonshot AI, MIT). Tested on one RTX PRO 6000 Blackwell.

    A pruned, quantized, and briefly retrained copy of Moonlight-16B-A3B-Instruct.

    6.7 GB

    from about 32 GB.

    427 / 500

    on fresh held-out math problems. The original got 432.

    Weak spot: coding. Lost all 3 coding prompts in a blind check.

    Weights on Hugging Face Read the thread

More releases as they're ready.

Small circular crop of the ColdTensor figure mark.

"Advanced AI shouldn't need a data center. Open research on making strong models run anywhere, with every tradeoff shown."

FLOE-01 weights inherit the MIT license from the base model.