36,151,104,512 per checkpoint (36B dense; Psyche and centralized variants)
512K (524,288 tokens advertised/configured; centralized training preset uses 131,072-token sequences)
Apache 2.0 declared for Psyche BF16 and official GGUF; centralized derivative license not explicitly stated. Upstream Seed Apache 2.0 notice/redistribution obligations apply; standalone derivative LICENSE files absent.
Text-only instruction/reasoning finetunes of ByteDance-Seed/Seed-OSS-36B-Base, not Hermes 4's Llama checkpoints. Same launch offers the production Psyche-trained BF16 model, an independently centrally trained BF16 research artifact, and official GGUF conversions of the Psyche model. Each BF16 checkpoint has 64 dense Transformer layers, 5,120 hidden width, grouped-query attention with 80 query/8 KV heads, untied embeddings and no exported vision or MTP module. Nous calls this its first production model post-trained entirely on Psyche: TP+DisTrO distributes training over the internet, with Solana coordinating consensus and a separate P2P gradient network; the comparison run uses its TorchTitan fork with FSDP+AdamW. The publisher reports 144K training tokens/second across 24 Psyche nodes and a run twice as large as Hermes 4, not foundation pretraining from scratch. Its stated Hermes 4 recipe is supervised instruction/reasoning training using verified synthetic trajectories; Atropos rejection sampling is data generation, not evidence of DPO/GRPO or on-policy RL on the released policy. Exact 4.3 corpus and processed-token totals are not established by the copied Hermes 4 marketing figures. The Seed base already configures 512K, so this is not a verified 128K-to-512K extension of that parent. Llama-style role headers are actually used despite the Seed architecture. The template defaults thinking=False and keep_cots=False; thinking=True selects a reasoning system prompt only when the user has not supplied the first system message. keep_cots controls historical reasoning retention, not a hard decoding budget. Tools are serialized inside <tools>, calls inside <tool_call>, and replies inside <tool_response>; structured JSON and reduced refusal are training goals. Publisher evaluations for the Psyche checkpoint report AIME 2024 71.9% and AIME 2025 69.3% estimated pass@1 from 64 samples, GPQA Diamond 65.5% from eight samples, MATH-500 93.8% from four samples, and IFEval 77.9% instruction-level loose accuracy (65.4% prompt-level strict). This is not best-of-64 or a tool-agent success rate. It exceeds the centralized 36B comparison across the published suite but does not uniformly exceed Hermes 4 70B: SimpleQA fuzzy-match is 6.0% versus 17.9%, and several reasoning/knowledge scores are lower. RefusalBench reports 74.60% non-reasoning and 72.29% reasoning, averaged over five trials; the claimed non-abliterated-model SOTA measures refusal behavior, not universal correctness or safety. Evaluation logs also report 3.4% of samples missing a closing thinking tag, including 18.8% on AIME 2025; exact server-side output limits and tool/retry settings are not fully specified. Quantized GGUF deployment is distinct from the approximately 72.3GB BF16 weight payload and does not guarantee consumer-GPU fit at full 512K context. December 3 is the publisher-declared public release/download-offer date; exact first public unlock remains unresolved. The centralized package lacks an explicit derivative license declaration, and both BF16 packages retain Seed EOS id 2 while the tokenizer's chat EOS is id 155127, requiring serving/stop-token care.