Field Guide - Model Registry

Using the Filters

At the top of the registry, you'll see filter buttons that help you narrow down models. Click any filter to activate it—click again to deactivate.

Capability Filters

Open Weights

Models whose weights can be downloaded and run on your own infrastructure. Important for reproducibility and long-term access.

Vision

Models that can "see" images—needed if you want to work with charts, screenshots, photos, or scanned documents.

Tool Use

Models that can call external tools—like searching databases, fetching citations, or running calculations.

Structured Output

Models that can reliably produce formatted data like JSON or structured tables. Useful for data extraction tasks.

Long Context

Models that can handle 100K+ tokens at once—great for long documents, comparing multiple sources, or extended conversations.

License Filters

Permissive

Models with minimal legal restrictions (MIT, Apache-2.0). Best for reproducibility and institutional use.

No Clickthrough

Models you can access without agreeing to special terms. Simpler for institutional deployment.

Tip: You can combine multiple filters. For example, click both "Vision" and "Permissive" to find open-licensed models that can process images.

Reading a Model Card

Each model appears as a card. Here's what the different parts mean.

The Header

Model Name — The name you'll use to identify the model (e.g., "Gemma 3 27B").

Vendor — Who created it (e.g., "Google", "Meta", "OpenAI").

Access Badges (top-right corner)

Open

Weights are freely downloadable without agreeing to special terms.

Gated

You'll need to agree to terms or request access before downloading.

Capability Badges

Colored pills that show what the model can do:

Vision Tool use Long context Open weights Reasoning mode

These badges give you a quick visual scan of capabilities without clicking into details.

The Specs Row

Below the badges, you'll see a gray bar with technical specifications. Here's what each number means.

70B / 16B active

This shows the model's size in billions of parameters.

70B — The total size. Bigger models often handle complex tasks better but are slower.
/ 16B active — If shown, this model uses a "mixture of experts" design. Only 16B parameters activate per request, making it faster while still being capable.

128K → 200K

This shows the context window—how much text the model can consider at once.

128K — The standard context size (~96,000 words).
→ 200K — If shown, the model can extend to this larger size with special configuration.

Rule of thumb: 1,000 tokens ≈ 750 words. So 128K tokens ≈ 96,000 words (a medium-length novel).

MoE or Dense

The model's internal design:

Dense — All parts of the model work on every request. Simpler and more predictable.
MoE — "Mixture of Experts." Only some parts activate per request, giving you big-model quality at lower cost.

Structured Output Status

Native — Model reliably produces JSON/structured data.

Prompted — Can produce structured output if you ask carefully.

Unknown — Not confirmed.

Understanding Licenses

Each card shows a license bar indicating what you can do with the model.

License Family Badges

The colored badge tells you the general category of restrictions:

Permissive

Minimal restrictions (MIT, Apache-2.0)

Use freely for research and teaching. Best for reproducibility.

Permissive+

Modified permissive (Modified MIT)

Mostly open with some extra requirements like attribution.

Community

Open with usage policies (Llama-style)

Free to use but with acceptable use policies and sometimes commercial thresholds.

Restricted

Terms of use required (Gemma-style)

Must agree to terms. May have use restrictions.

License ID & View Link

Next to the badge, you'll see the specific license name (e.g., "MIT", "Apache-2.0"). Click View to read the full license text.

For institutional use: If you need models for long-term reproducible research, prioritize Permissive licenses and models marked Open. Always consult your institution's guidance for licensing decisions.

Expanded Details

Click "Show details" at the bottom of any card to see more information.

Best For

A list of tasks this model excels at—like "long-document reading," "coding assistance," or "multilingual work."

+ Strengths

What the model does well. Examples: "strong at staying coherent with very large context," "good multilingual coverage."

– Limitations

What to be aware of. Examples: "no image input," "tool-calling depends on provider."

License Notes & Context Notes

Additional details about licensing terms or context window behavior. Not all models have these.

Choosing a Model

Quick guidance based on what you need to do.

Long Documents

Reading long texts, comparing sources, extended conversations

Look for: Long context badge, 100K+ context

Images & Charts

Screenshots, diagrams, scanned documents, photos

Required: Vision badge—no workarounds

Code & Data

Writing code, extracting data, working with structured formats

Look for: Tool use, Native structured output

Reproducibility

Long-term research, citable outputs, institutional use

Look for: Permissive license + Open badge

A note on "accuracy"

Bigger or "reasoning" models aren't automatically more accurate. For scholarly work:

• Always verify against original sources
• Ask for citations and check them
• Treat models as drafting tools, not authorities

Glossary

Terms you might see in the registry.

Token: A chunk of text the model processes—roughly 3-4 characters or ¾ of a word.
Context: How much text the model can "see" at once. 128K context ≈ 96,000 words.
Parameters (B): The model's size in billions. Bigger often means more capable but slower.
Dense: A model where all parameters work on every request.
MoE: "Mixture of Experts"—only some parameters activate per request, balancing capability and speed.
Open Weights: The model can be downloaded and run on your own servers.
Tool Use: The model can call external tools like search, calculators, or databases.
Structured Output: The model can produce formatted data like JSON reliably.

Back to Model Registry

How to Use the Model Registry