Thinker: nominal 30B total / 3B active; full Instruct checkpoint: 35,259,818,545 stored parameters; Thinking: 31,719,205,488; Instruct adds a nominal 3B / 0.3B-active Talker and speech-generation modules
65,536 tokens configured for Thinker; Instruct Talker also 65,536 (code predictor: 32,768)
Apache 2.0 (explicit checkpoint/model-family grant in the technical report; standalone LICENSE files absent from the three HF repositories)
Native multimodal Thinker–Talker MoE family accepting text, images, audio and video, including temporally aligned video audio. Instruct produces text and streaming speech; the separate Thinking checkpoint contains only the Thinker and produces chain-of-thought text, not speech. Select the checkpoint explicitly rather than assuming a shared thinking-mode toggle. The Thinker has 48 layers, 128 routed experts with eight selected per token, GQA and interleaved multimodal rotary positions; the Talker uses 128 routed experts with six selected plus a shared expert. AuT audio and SigLIP2-initialized vision encoders supplement the language decoder, so 30B/3B does not describe the whole speech pipeline. Speech generation uses 16 discrete codec codebooks, residual-code prediction and a lightweight causal Code2Wav ConvNet instead of block-wise diffusion, with 12.5-Hz codec frames. The report describes 119 text languages, 19 speech-input languages and 10 speech-output languages. Original Instruct exposes three voices: Chelsie, Ethan and Aiden. Publisher-reported theoretical cold-start first-packet latency is 234 ms for audio-only and 547 ms for audiovisual input at concurrency one with optimized vLLM; this is not a guaranteed network-inclusive response time. Pretraining aligns encoders/adapters to a frozen Qwen3-initialized LLM, then jointly trains on roughly 2T multimodal tokens and extends training sequences from 8,192 to 32,768. AuT was separately trained on 20M hours of supervised/pseudo-labeled audio. Thinker post-training combines SFT, off-/on-policy teacher distillation and GSPO with rule/model rewards; Talker uses multimodal speech training, high-quality continued pretraining, multilingual DPO and speaker fine-tuning. Training sequence lengths and reported 30–40-minute audio handling are not checkpoint context-window guarantees. In publisher evaluations, AIME25 is 65.0 for Instruct and 73.7 for Thinking, while BFCL-v3 is 64.4 and 63.2 respectively; evaluation uses greedy Instruct decoding, checkpoint sampling settings for Thinking, no system prompt and two video frames per second. The headline 32/36 open-source and 22/36 overall audio/audiovisual SOTA claim covers the evaluated suite including unreleased in-house Flash variants, not a guarantee for every downloadable checkpoint. The report identifies weaker long-video performance due to context and positional-extrapolation limits. Related same-launch Captioner is an Instruct-derived, single-turn audio-only/text-output fine-tune accepting no user text prompts, with clips of at most 30 seconds recommended; it is referenced but excluded from this text-prompt-capable grouping. Hosted Flash, experimental Base comparisons and later 2512 checkpoints are not included. September 22 is verified availability-by from the explicit public release/download materials and contemporary open-weight reporting; the exact first public unlock remains unresolved.