You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix vignette layout: replace Quarto ::: divs with inline HTML flex containers
Quarto's layout classes (quarto-layout-row, quarto-layout-cell) depend on
CSS in vignette_files/ which is gitignored and not served by htmlpreview.
Replacing all five image layout blocks with inline-styled <div> flex containers
makes the side-by-side screenshots render correctly without external CSS.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
<li>Under <strong>Ranges</strong>, drag the <strong>Size (B parameters)</strong> slider to a maximum of 8.60 B</li>
173
173
</ol>
174
174
<p>The grid immediately narrows to twelve models, all with an openness score of 5: Apertus 8B, GPT-J 6B, OLMo 2 7B, OLMo 3 7B, OLMo 7B, and seven Pythia models (Pythia 1.4B, Pythia 160M, Pythia 1B, Pythia 2.8B, Pythia 410M, Pythia 6.9B, and Pythia 70M). Click OLMo 2 7B to open its profile card, which shows the foundational paper link and fetches the three most recent arXiv papers mentioning the model. If your GPU budget extends to 32 B, also remove the size cap — OLMo 2 32B and OLMo 3 32B pass every other criterion, and OLMo 3 32B also has a think variant available.</p>
<li>Open <strong>Exclusion filters</strong> → <strong>Exclude country of origin</strong> → select or type <strong>United States</strong> and <strong>China</strong></li>
215
191
</ol>
216
192
<p>Eight models remain: Apertus 8B (Switzerland, score 5), BLOOM 176B and the three Mistral models (France), the two Falcon models (United Arab Emirates), and Sarvam 30B (India, score 2). The scatter plot makes the trade-offs immediately visible: Apertus 8B leads on openness; BLOOM covers 46 languages but carries a BigScience RAIL license with restrictions on certain uses; the two Falcon models cover 4 languages each but are Apache 2.0 licensed.</p>
<p><strong>Step 2 — Filter by a specific language.</strong> Use this independently — without the openness filters active — when you need to know which models officially support a particular language, regardless of their openness profile.</p>
<li>Move the <strong>Context window (tokens)</strong> slider minimum to <strong>8,192</strong></li>
329
251
</ol>
330
252
<p>Eight models appear. On the scatter plot with context window on the y-axis, three models stand out at 131 K context — Qwen2 7B, Qwen2.5 7B, and Llama 3.1 8B. At 65,536 tokens, Apertus 8B and OLMo 3 7B occupy the middle tier — both Apache 2.0 licensed with an openness score of 5 and OLMo 3 7B additionally carrying a think variant. Qwen3 8B sits at 32,768 tokens. Mistral 7B and Gemma 2B complete the results at 8,192 tokens. Click each model to compare licenses and training details in the profile card before deciding.</p>
<li>Under <strong>Openness filters</strong> → check <strong>Open weights</strong>, <strong>Open training data</strong>, <strong>Intermediate checkpoints</strong>, <strong>Open code</strong>, and <strong>Permissive license</strong></li>
365
275
</ol>
366
276
<p>Fifteen models score 5: Apertus 8B, GPT-J 6B, OLMo 2 7B, OLMo 2 32B, OLMo 3 7B, OLMo 3 32B, OLMo 7B, and all eight Pythia models. To narrow to fully open models with a think variant, also check the <strong>Think version available</strong> checkbox — two models remain: OLMo 3 7B and OLMo 3 32B. Click <strong>OLMo 3 7B</strong> to open its profile card — it and OLMo 3 32B are the only fully open models in the database with a chain-of-thought (think) variant. OLMo 3 7B is the recommended starting point for most research workflows — it is more accessible than the 32B variant due to its lower compute requirements. Check the recent arXiv papers fetched in the profile card to confirm that OLMo 3 7B is being actively used in the literature and that there are relevant papers building on it.</p>
Copy file name to clipboardExpand all lines: docs/vignette.qmd
+55-80Lines changed: 55 additions & 80 deletions
Original file line number
Diff line number
Diff line change
@@ -77,21 +77,16 @@ In the sidebar:
77
77
78
78
The grid immediately narrows to twelve models, all with an openness score of 5: Apertus 8B, GPT-J 6B, OLMo 2 7B, OLMo 3 7B, OLMo 7B, and seven Pythia models (Pythia 1.4B, Pythia 160M, Pythia 1B, Pythia 2.8B, Pythia 410M, Pythia 6.9B, and Pythia 70M). Click OLMo 2 7B to open its profile card, which shows the foundational paper link and fetches the three most recent arXiv papers mentioning the model. If your GPU budget extends to 32 B, also remove the size cap — OLMo 2 32B and OLMo 3 32B pass every other criterion, and OLMo 3 32B also has a think variant available.
Eight models remain: Apertus 8B (Switzerland, score 5), BLOOM 176B and the three Mistral models (France), the two Falcon models (United Arab Emirates), and Sarvam 30B (India, score 2). The scatter plot makes the trade-offs immediately visible: Apertus 8B leads on openness; BLOOM covers 46 languages but carries a BigScience RAIL license with restrictions on certain uses; the two Falcon models cover 4 languages each but are Apache 2.0 licensed.
**Step 2 — Filter by a specific language.** Use this independently — without the openness filters active — when you need to know which models officially support a particular language, regardless of their openness profile.
154
142
@@ -158,21 +146,16 @@ In the sidebar (with all other filters reset):
Eight models appear. On the scatter plot with context window on the y-axis, three models stand out at 131 K context — Qwen2 7B, Qwen2.5 7B, and Llama 3.1 8B. At 65,536 tokens, Apertus 8B and OLMo 3 7B occupy the middle tier — both Apache 2.0 licensed with an openness score of 5 and OLMo 3 7B additionally carrying a think variant. Qwen3 8B sits at 32,768 tokens. Mistral 7B and Gemma 2B complete the results at 8,192 tokens. Click each model to compare licenses and training details in the profile card before deciding.
Fifteen models score 5: Apertus 8B, GPT-J 6B, OLMo 2 7B, OLMo 2 32B, OLMo 3 7B, OLMo 3 32B, OLMo 7B, and all eight Pythia models. To narrow to fully open models with a think variant, also check the **Think version available** checkbox — two models remain: OLMo 3 7B and OLMo 3 32B. Click **OLMo 3 7B** to open its profile card — it and OLMo 3 32B are the only fully open models in the database with a chain-of-thought (think) variant. OLMo 3 7B is the recommended starting point for most research workflows — it is more accessible than the 32B variant due to its lower compute requirements. Check the recent arXiv papers fetched in the profile card to confirm that OLMo 3 7B is being actively used in the literature and that there are relevant papers building on it.
0 commit comments