Canonical: https://northloom.ai/technical-work/

Technical workMeasured, then explained

# Results before adjectives.

We publish measured work so you can see what the products do and how we think. The first set covers large-model inference with Krasis.

Krasis / Benchmarks

## Large models on single-GPU systems.

These results show three models running with Krasis. They are individual demonstrations on different systems, not a direct ranking between models or graphics cards.

01Measured result

## Ornith 397B

1× Nvidia RTX PRO 6000

96 GB VRAM · 200 GB DDR4

**2,354**tokens/secPrefill

**25.72**tokens/secDecode

02Measured result

## Step 3.7

1× Nvidia RTX PRO 6000

Single-GPU system

**5,261**tokens/secPrefill

**55.4**tokens/secDecode

03Measured result

## Qwen 3.5 122B

1× Nvidia RTX 5090

Single-GPU system

**3,692**tokens/secPrefill

**31.8**tokens/secDecode

Reading the results

01

## Prefill

Prefill is the work of reading and processing the input before a response begins. Faster prefill matters when prompts and context are large.

02

## Decode

Decode is the generation of the response itself. Its token rate is the speed a person experiences once the model starts answering.

03

## Reproducibility

Model configuration, prompt length and software versions affect results. Full test details will be published alongside future engineering notes.

Your system

## Unsure where the bottleneck is?

Northloom offers focused, remote reviews of self-hosted inference systems and proposed hardware designs.

[Explore consultancy services ↗](https://northloom.ai/consultancy/)
