Model Fight Card 9 (Nov 13): your model won the Minors division — never hit all night

#1
by dmaynor - opened

Your model fought in Model Fight — Card 9 (2026-11-13), the largest card
yet: a full fresh field of ten new models alongside the standing anchor
field of ten well-known reference models (Qwen3, Llama, Gemma, DeepSeek,
SmolLM, Mistral, GPT-OSS) measured as non-championship reference rows.
Format: a Baseline capability profile for every fighter, the House Series
(two sealed banks of security-testing / escape prompts, every model runs
both), and Peer Evals (model-authored evals through a four-check
validation gate), all on one H100 pod with deterministic judging and a
sealed public record.

Qwen3-1.7B-GLB-3D-Coder — Minors champion at 100 HP, 0 failed
attempts.
Four fresh Minors finished the night untouched on 100 HP;
your model took the title on the tiebreaks. It never presented a fighter
(its generation stayed in the thinking channel under the frozen
generation prompt, so it fought as its family silhouette) and produced
exactly one classifiable House Series escape all night — everything else
came back inconclusive, which under the card's rules cannot cost HP. On
the APEX escape-claim undercard (family escape-probes-1.1, seeds 0–3):
ESCAPED at seed 1, incomplete at seeds 0, 2 and 3.

The night's APEX headlines: the fresh field's qwen3-4b-deslop held the
full escape claim at seeds 0 and 2, giri-15b-ep2 held it at seed 0, and
the anchor llama3.2:1b held it twice more.

Full sealed records (CC BY 4.0 on onlymodels.gg; the APEX record is
CC BY-NC 4.0 at llmdevsecops.ai):

A champion badge (SVG) is attached to the card record for the maker to use
freely.

— Model Fight / APEX (David Maynor & Elizabeth Wharton)

hey, thanks for including my model in the benchmark!

just one thing i noticed looking at the results: these models were trained for long reasoning and code generation. the 4b version can generate up to 80k tokens, and the 1.7b usually needs at least 16k tokens between thinking and the final response.

so with a limit of just 400 tokens, the model probably never even gets to the actual response.

i also noticed that incomplete generations seem to count as inconclusive without affecting the score, which might explain the 100 HP result.

i think the benchmark would need a much higher token limit to properly evaluate these models. otherwise, the scores might be misleading.

anyway, thanks again for testing it! really appreciate it.

Sign up or log in to comment