This is a full-parameter BF16 supervised fine-tune of Qwen/Qwen3-30B-A3B-Base for long-form mathematical reasoning.

The model was trained on 500,379,618 supervised assistant tokens from a 4K-12K-token chain-of-thought subset of nvidia/Nemotron-SFT-Math-v4. It uses the Qwen3-30B-A3B-Thinking-2507 tokenizer, chat template, and generation configuration so inference begins in a <think> block.

User tokens, the assistant header, the opening <think> prefill, and pack padding were masked from the loss. Supervision begins at the first reasoning token and includes the reasoning trace, </think>, final answer, and <|im_end|>.

Evaluation used the 30 problems in math-ai/aime25, revision 563bb8404243c5f09de6ec262f2db674fe5bce9b, with 16 sampled responses per problem (480 responses total).

Sampling used temperature 0.6, top-p 0.95, top-k 20, BF16 weights, and a BF16 KV cache. Each prompt requested step-by-step reasoning and a final answer in \boxed{}.

Avg@16 is the fraction of all 480 sampled responses that are correct. Pass@16 is the empirical fraction of problems solved by at least one of their 16 responses; it is not an estimator computed from another sample count.

The 32,768-token result is the native generation run. The 8,192- and 16,384-token rows were obtained by truncating and rescoring those same sampled trajectories at the corresponding exact prompt plus completion boundaries. This gives a paired context-length comparison rather than three independently sampled evaluations. Scores were unchanged when each truncation boundary was perturbed by plus or minus 64 re-tokenized tokens.

For reference, a separate independently sampled run with a 16,384-token response cap (and 17,408-token model context) scored 16.04% Avg@16 and 43.33% Pass@16; it is not directly comparable to the exact-total-context table.

At the 32,768-token total-context limit, 129/480 responses (26.88%) stopped naturally and 351/480 (73.13%) reached the context limit. Naturally stopped responses were correct 87/129 times, while context-limited responses were correct 1/351 times. The checkpoint therefore still has a substantial overlong-reasoning and termination failure mode.

This checkpoint is intended as a research starting point for mathematical reasoning post-training. It is not guaranteed to produce correct proofs or to terminate before the context limit. Generated reasoning and final answers should be independently verified.