30.532B total; approximately 3.353B active per token (MoE, including input/output embeddings)
262,144 tokens (256K native); 1M–4M document processing requires iterative memory-agent inference
Apache 2.0 (publisher-declared; standalone checkpoint LICENSE files absent)
Text-only reasoning checkpoint post-trained from Qwen3-30B-A3B-Thinking-2507, with 48 Transformer layers, grouped-query attention (32 query/4 KV heads), and 128 routed experts selecting eight per token; released in BF16. Long-context data synthesis builds multi-hop questions from atomic document facts and globally distributed evidence, including corpus-level numerical reasoning and dialogue memory; 14.1K training examples survive filtering from 42.7K synthesized examples. Purely on-policy, token-normalized GRPO uses task-balanced sampling, task-specific reward-standard-deviation normalization, and Adaptive Entropy-Controlled Policy Optimization (AEPO), which switches negative-advantage masking on and off at upper/lower entropy thresholds. Rewards combine ground-truth containment checks with gpt-oss-120b judging when rule checks fail. Three full-context RL stages increase input/output budgets from 20K/12K to 60K/20K and 120K/50K; a separately trained memory-management expert is SCE-merged with Stage 3, followed by a fourth full-context RL stage. This produces one released checkpoint, not separate full-context and memory models: the memory agent repeatedly updates compact memory and a navigational plan across document chunks rather than expanding native attention to millions of tokens. The report averages 71.82 across six long-context benchmarks versus 61.92 for its parent, with MRCR 82.99 and CorpusQA 81.25; these evaluations cap input at 128K and generation at 50K, use middle truncation, and combine answer matching with model judging for QA. The abstract claims a 9.48-point ultra-long-task gain over the parent memory agent, with aggregation unspecified; Table 9 reports only 14.29 on CorpusQA at 4M tokens. LongMemEval rises from 60.80 to 76.40 and AIME25 from 82.81 to 86.46, while AIME24 and BFCL-V4 Memory-Rec_Sum slightly regress. Multilingual capability is inherited, but predominantly English post-training may weaken non-English long-context reasoning. Authors identify long-form-output coverage and subjective-task reward quality as limitations; the training dataset was not publicly released.