abacusai/Smaug-Flash

Figures: LiveBench Mathematics 90.1 (was 91.3), overall 77.4 / +3.2, matching the 2026-06-25 leaderboard

Update benchmark figure: AutomationBench 38.8 (+13.7)

Update AutomationBench delta to +13.7 (38.8 vs 25.1)

Update README.md

Update README.md

Update README.md

Remove unselected figure options

Finalize evaluation: keep grouped bar chart

card fixes: gen-config top_p 0.95 (match recommended sampling), Hidden Dimension label, three-stage LoRA row, sglang kv-cache-dtype fp8_e4m3, recipe URL -0731

Drop 'agentic cells' jargon; normalize DeepSeek harness casing

AA-LCR noise caveat stated once (Notes only)

Fix caption/notes contradiction on starred rows, parity wording, SGLang flag mismatch

Drop section cross-reference in evaluation intro

Delta pill corner radius matched to bar chart

Remove superseded single vs-base figure

Three vs-base figure options (bars/dumbbell/table); radar delta bars aligned with table style

Terminal Bench 2.1 DeepSeek-harness score updated to 76.4 (+10.1); wrap harness labels

Radar axis labels clear of the circle

Trim outer canvas padding on figures (~5% larger render)

Delta bars restored with doc-matched geometry; doc-matched column and chart-table spacing

Agentic-style structure: model summary, eval as section 3, delta values only, short eval notes

Delta bars: rectangles across centered zero-line, matching the doc

Doc-style delta bars in radar; reorder sections to Mini structure; condense training detail

Add vs-base and LiveBench category charts, Abacus header, tags, citation; replace eval table

TB2.1 terminus-2: 70.8 (2026-09-02 remeasure, interleaved thinking; 69.7 at uniform 1x)

DeepSWE row: bold 56.6, drop redundant footnote clause

DeepSWE row: simplify label, base = vendor-card 54.4*

eval: DeepSWE v1.1 official-harness score 56.6 (2026-09-02 full 113-task run)

Card cleanup: replace inherited DeepSeek header with Smaug-Flash lead, add base_model metadata, fold usage/serving docs, add AA-LCR paired result, license/citation/contact hygiene

Finalize DeepSWE row: paired parity (49.1 vs 50.9, n=109, p=1.0)

Add technical report: training methodology, data curation, weight merging, evaluation

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Upload README.md with huggingface_hub

initial commit