About this map
A method is only listed here if a source describes it in a sentence that can be quoted. Extraction reads each document, names the methods it substantively treats, and files each one into a hand-curated taxonomy; a path the model invents is rejected rather than accepted, and shows up below as a proposal for a human to promote.
1 733
techniques
2 955
evidence spans
231
unfiled
no taxonomy path
3
taxonomy proposals
awaiting review
How a page here is produced
- Ranking. OpenRouter's weekly open-weight token ranking is refreshed automatically; new entrants land in a review queue.
- Curation. A human adds the model's documents — blog post, technical report, model card, repository — to a checked-in source list.
- Fetch. Each document is fetched and reduced to text. The curation labels (which model, which publisher) are withheld from the extractor, so what it finds is not what it was told.
- Extraction. Per document: what kind of document it is, then every method it substantively treats, each with a required evidence quote, an adoption status, and the artifact it applies to.
- Filing. Methods are grouped, filed into the curated taxonomy, and merged into named techniques. Nothing invents a taxonomy path.
- Code checks. Separately, each model’s modeling code and config.json — from Hugging Face Transformers, the model repository, or vLLM, pinned to a commit — are checked against a short, curated list of architecture features. The verifier is not told the model’s name or what its documents say, and a verdict is shown only with the config values and source lines it cites that were found in the files.
Who maintains this
I'm Yannick Versley, and I maintain this map because open-weight models are a positive force — for education, and for the self-determined, flexible use of large language models outside a handful of commercial APIs. Knowing more about what actually happens inside these models helps clarify the fundamental problems they still have to solve.
Code checks
19 features checked in the code of 28 models.
Run
| Stage | Prompt version |
|---|---|
| classification | radar-classification-v2 |
| concepts | radar-concept-v1 |
| entities | radar-entity-v1 |
| events | radar-event-v0 |
| methods | radar-method-v1 |
| proposals | radar-proposal-v1 |
Model: openai/gpt-6-luna · run 2026-09-25T22:33:20 · 169 documents · {'calls': 122, 'rate_limited': 0, 'retries': 0} LLM calls.
Taxonomy proposals
What the corpus wanted and the curated outline does not have. These are not published as techniques until a human promotes them.
| Parent | Proposed | Rationale |
|---|---|---|
| model architecture :: token mixer :: softmax attention | gated attention | Attention outputs are modulated by a learned gate as part of the attention mechanism. The existing grouped-query attention node describes how query and KV heads are grouped, not ga |
| model architecture :: hybrid layer stacking | linear-attention and sparse-attention hybrid | These methods interleave stateful Gated DeltaNet layers, which compress history, with sparse-attention layers that retrieve selected long-range context. The existing hybrid stackin |
| data curation | training data format | These groups describe structured serialization of training and inference conversations, including message boundaries, channels, and encoding into model input strings. Tokenization |
Known gaps
45 failures in the last run — documents or stages that did not complete. Listed because a missing method is invisible otherwise.
| Where | What happened |
|---|---|
| concepts | Expecting ',' delimiter: line 20 column 15 (char 588) |
| concepts | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | 1 validation error for MethodExtraction item_id Field required [type=missing, input_value={'': None}, input_type=dict] For further information v |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| events | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| events | 5 validation errors for EventDetection item_id Field required [type=missing, input_value={'events': [{'event_type'...is the model release. |
| classify | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| classify | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| classify | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| methods | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| methods | Expecting ':' delimiter: line 160 column 23 (char 8011) |
| methods | LLM returned an empty response (a reasoning model can spend the whole output budget on reasoning tokens; try --reasoning none) |
| classify | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| concepts | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| propose | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
| entities | LLM response hit the 32768-token output cap (raise --max-output-tokens / --max-normalization-output-tokens, or lower --method-chunk-chars / --normalization-max- |
Data sources
Model ranking: Source: OpenRouter (openrouter.ai/rankings), as of 2026-09-25T22:22:06.301Z. Licensed CC BY 4.0.
Document text belongs to its publishers; quotes here are short excerpts shown with attribution and a link to the original.