OLMo-2-7B
7.3B
K#
141.1
A
Top-tier structural quality at workhorse size. Fully open weights and training data from Allen AI.
Download Report↓AI · Product
A structural report on any transformer model. Twelve measurements you can act on, delivered as a PDF you can read in an afternoon. No training, no fine-tuning, no benchmark theater.
Standard tier · $1,500 per report · 72 hours from commencement
Request a Keel Scan See tiersWhat it measures
Every Keel Scan report is a structural read on the model as-shipped. These twelve outputs appear in every Standard report.
A single number on a 0–200 scale. 100 is the Keel Index baseline. Higher is better structure.
A letter A through F. Fast, comparable, calibrated against every model on the Keel Index.
The shape of the model. Sequential, concentrated, front-loaded, declining, or dual-phase.
How well the layers of the model work together. High coherence, low internal friction.
How much of the model is doing useful work versus running hot. Efficiency, in a word.
Where this model sits on the capability ladder. Compared like-for-like against peers.
How the model activates across its layers. The shape reveals what is doing the work.
A per-layer read. Which layers earn their place. Which layers are passengers.
What happens if you remove specific layers. Which are essential, contributing, or redundant.
How much of the parameter budget is actually representing knowledge.
Where this model sits against every other model in its size class on the Keel Index.
Concrete actions ranked by impact. What to prune, what to quantize, what to leave alone.
Featured samples
Real reports on real models, drawn from the Keel Index. Each is a full 19-page structural read, formatted exactly as you would receive one.
OLMo-2-7B
7.3B
K#
141.1
Top-tier structural quality at workhorse size. Fully open weights and training data from Allen AI.
Download Report↓Llama-3.1-8B
8.0B
K#
136.0
Meta’s widely-deployed 8B foundation model. Common base for fine-tunes across the ecosystem.
Download Report↓Qwen3-32B
32.8B
K#
134.0
Latest generation Qwen. Sits next to Qwen2.5-32B on the Index below, at a lower structural score.
Download Report↓gpt2
124M
K#
92.8
The 2019 model that started the modern era. Small, historically important, structurally simple.
Download Report↓The Keel Index
Thirty-four transformer models scanned so far, ordered from highest K# down. The Index is Keel Technologies’ public reference for structural quality across the open-model ecosystem.
| Model | Parameters | K# | Grade | Report |
|---|---|---|---|---|
| Qwen/Qwen2.5-32B | 32.8B | 153.6 | A | · |
| 01-ai/Yi-1.5-34B | 34.4B | 142.4 | A | · |
| allenai/OLMo-2-1124-7B | 7.3B | 141.1 | A | PDF↓ |
| mistralai/Mistral-7B-v0.3 | 7.2B | 140.1 | A | · |
| Qwen/Qwen2.5-7B | 7.6B | 137.4 | B | · |
| meta-llama/Llama-3.1-8B | 8.0B | 136.0 | B | PDF↓ |
| microsoft/phi-4 | 14.7B | 135.4 | B | · |
| Qwen/Qwen2.5-Coder-7B | 7.6B | 134.4 | B | · |
| Qwen/Qwen3-32B | 32.8B | 134.0 | B | PDF↓ |
| bigcode/starcoder2-7b | 7.2B | 133.9 | B | · |
| google/gemma-2-9b | 9.2B | 132.5 | B | · |
| meta-llama/Llama-3.1-8B-Instruct | 8.0B | 131.2 | B | · |
| meta-llama/Llama-3.2-3B | 3.2B | 129.3 | B | · |
| EleutherAI/gpt-j-6b | 6.1B | 128.5 | B | · |
| Qwen/Qwen2.5-1.5B | 1.5B | 128.5 | B | · |
| Qwen/Qwen3-14B | 14.8B | 128.0 | B | · |
| Qwen/Qwen2.5-1.5B-Instruct | 1.5B | 127.7 | B | · |
| EleutherAI/pythia-6.9b | 6.9B | 127.6 | B | · |
| google/gemma-2-2b | 2.6B | 126.6 | B | · |
| Qwen/Qwen2.5-Math-7B | 7.6B | 125.5 | B | · |
| HuggingFaceTB/SmolLM2-360M | 362M | 125.4 | B | · |
| Qwen/Qwen3-8B | 8.2B | 124.5 | B | · |
| meta-llama/Llama-3.2-1B | 1.2B | 123.6 | B | · |
| TinyLlama/TinyLlama_v1.1 | 1.1B | 122.4 | B | · |
| EleutherAI/pythia-1.4b | 1.4B | 121.5 | B | · |
| Qwen/Qwen3-4B | 4.0B | 117.0 | B | · |
| facebook/opt-1.3b | 1.3B | 116.6 | B | · |
| Qwen/Qwen3-1.7B | 1.7B | 111.2 | B | · |
| deepseek-ai/DeepSeek-R1-Distill-Llama-8B | 8.0B | 110.3 | B | · |
| Qwen/Qwen3-0.6B | 596M | 105.2 | B | · |
| deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | 7.6B | 103.6 | B | · |
| google/gemma-3-1b-it | 1.0B | 97.9 | C | · |
| gpt2 | 124M | 92.8 | C | PDF↓ |
| distilgpt2 | 82M | 83.0 | C | · |
Tiers
Every tier delivers the same twelve measurements. Higher tiers add capacity, scoping, and the bespoke treatment larger models require.
Standard
$1,500
Any dense transformer up to forty billion parameters on Hugging Face. The fastest way to get a Keel Scan on a model you can share.
Scan+
$5,000
Larger dense transformers, forty to seventy-five billion parameters. Includes a scoping conversation to tailor the analysis.
Frontier
from $25,000
Above seventy-five billion parameters, mixture of experts, or proprietary weights. Signed SOW, custom security, delivery walk-through.
Full commercial terms in the Service Schedule.
Fill in the intake form and Keel Technologies will confirm tier and next steps within one business day.
Request a Keel ScanFrequently asked
Any dense transformer up to forty billion parameters on Hugging Face qualifies for the Standard tier. Forty to seventy-five billion parameters and gated repositories go to Scan+. Larger models, mixture of experts, and proprietary weights go to Frontier.
No. Public Hugging Face models are simplest, but Keel Technologies also scans gated repositories with your authorization, and proprietary weights under signed SOW at the Frontier tier.
Standard: seventy-two hours from commencement. Scan+: five to seven business days. Frontier: three to five weeks depending on scope. Commencement is the point at which payment is received and required access has been provided.
Report the nonconformity within the acceptance window and Keel Technologies will correct it, re-perform the scan, or refund the fee. The specific mechanics are in the Terms and Conditions.
Active copies of your model are deleted from Keel Technologies’ systems within seven days of Report acceptance. Full data handling terms are in the Terms and Conditions, Section 7.
Yes, within the limits in the Terms and Conditions, Section 10. You can share the Report internally without restriction, use the K# score and grade in external communications, and cite Keel Technologies as the source.