This record is a coverage caveat in itself: the report behind it is the Qwen3.5-Omni Technical Report, and it never documents the Qwen3.5 text backbone that this record's gallery card names (397B-A17B). Neither '397B' nor 'A17B' appears anywhere in the text; the only scale statement is that the Omni model 'scales to hundreds of billions of parameters'. What the report does disclose about the Qwen3.5 family: a 250k-vocabulary byte-level BPE tokenizer; a backbone it names 'Hybrid MoE' — 'Hybrid-Attention Mixture-of-Experts' — introduced in Qwen3.5 and reused for both the Thinker and Talker of the Omni system; and the fact that this architecture 'includes the Gated Delta Net (GDN) module', a delta-rule-style linear-attention component credited with cutting KV-cache I/O in long-context inference. The text-only counterpart Qwen3.5-Plus-Instruct serves as the evaluation baseline (Table 4) and is said to be 'same-size' as the Omni models. Around that skeleton the report details the Omni variant: Thinker–Talker streaming with a SigLIP2 vision encoder, an AuT audio encoder trained on 40M hours of audio-text data, RVQ multi-codebook speech with an MTP module and Code2Wav renderer, ARIA's adaptive text-speech interleaving, explicit timestamp tokens replacing TMRoPE for temporal grounding, and a three-stage pre-training (encoder alignment, then a 32,768-token general stage on ~4T tokens, then a 262,144-token long-context stage) followed by specialist distillation, on-policy distillation, and interaction-aligned RL. Everything about the text model's own layers, width, expert routing, context length, and training remains undisclosed; per schema §3, the scale/context/training leaves above describe the report's flagship (Qwen3.5-Omni-Plus) where stated.
What the report claims is novel
Scaling the omnimodal Thinker–Talker architecture to hundreds of billions of parameters with 256k-token context, using Hybrid-Attention MoE (including the GDN module) for both Thinker and Talker to enable efficient long-sequence inference [Abstract; §1; §2.5]
ARIA (Adaptive Rate Interleave Alignment): a single-stream interleaving of text and speech tokens under an adaptive rate constraint, replacing dual-track generation to stabilize streaming speech synthesis and improve prosody with minimal latency impact [Abstract; §2.4]
Expanded multilingual speech: 113 languages/dialects for speech recognition and 36 for speech synthesis, with zero-shot voice customization from user-provided samples [§1; Table 3]
Controllable audio-visual captioning: script-level structured captions with timestamp annotation, character-relationship description, and automated scene segmentation [Abstract; §1]
Native omnimodal agentic behavior: autonomous WebSearch, complex FunctionCall, and Audio-Visual Vibe Coding — emergent direct code generation from audio-visual instructions [Abstract; §1]
Replacing TMRoPE temporal IDs with explicit text-string timestamps and contiguous cross-modality position numbering for robust long audio-video temporal perception [§2.3; §3]
Atlas assessment: The genuinely novel surface of this report sits entirely on the Omni side — ARIA's adaptive-rate single-stream text/speech interleaving, timestamp-string temporal encoding replacing TMRoPE, and multi-codebook MTP streaming — none of which concerns the text backbone this record represents. The text-side disclosure reduces to a name: 'Hybrid MoE' with a GDN module, i.e. a dense/MoE hybrid containing a gated delta-rule linear-attention component, consistent with the hybrid-attention trend already in the atlas (qwen-qwen3-next's hybrid backbone; moonshot-kimi-k3's KDA/MLA hybrid). Because the report never documents the text model's configuration, its novelty relative to Qwen3-Next-style or DeepSeek-style hybrids cannot be assessed from this source. The claim that the Omni model matches 'same-size' text counterparts on text benchmarks (§1, Table 4) is internal to the report and not independently verifiable.
Relation to the atlas
Lineage in
Thinker–Talker dual-component architecture (Thinker generates text; Talker generates streaming speech conditioned on Thinker representations) from Qwen2.5-Omni
RVQ-based speech representation (multi-codebook codec tokens) from Qwen3-Omni
Chunked pre-filling (encoders output temporal chunks to cut Time-To-First-Token) from Qwen3-Omni / Qwen2.5-Omni
TM-RoPE temporal position encoding (applied, then superseded by explicit text-string timestamps with contiguous cross-modality position numbering) from Qwen3-Omni
ChatML conversation format for the training corpus from OpenAI
LLM initialization and vision encoder adopted from the Qwen3.5 text family (S1 encoder-alignment stage initializes the Omni LLM with Qwen3.5 parameters) from Qwen3.5 (text family)
Influence out
No descendants recorded yet.
Notable omissions
CRITICAL COVERAGE GAP: the report documents the Qwen3.5-Omni multimodal variant only — the Qwen3.5 text backbone (the 397B-A17B model of the gallery card) is never described; '397B'/'A17B' never appear, and no layers, hidden dim, heads, expert counts, activation, normalization, context length, or training details are given for the text model
No configuration table for any variant (no total/active parameters, layers, dims, heads, vocab aside from the 250k tokenizer, context)
Only scale statement for any model: 'hundreds of billions of parameters' (Abstract, Omni); the Plus vs Flash sizes are contrasted ('substantial scale difference', Table 2) but never quantified
Expert count, router, and load balancing of the MoE never described — 'Hybrid MoE' is a name, not a spec
GDN (Gated Delta Net) is named as part of the Qwen3.5 architecture but its mechanism is never explained
No optimizer, learning rate, batch size, precision, parallelism, hardware, or compute (FLOPs/GPU-hours) disclosed for any training stage
No statement on weight release or a model-weights license (CC BY 4.0 on the arXiv page is the paper license only)
Text model context length not disclosed (the 256k figure is stated for the Omni series); tokenizer disclosure limited to byte-level BPE + 250k vocab
No attention head/KV design (GQA/MLA), head dim, or embedding tying disclosed
Per-component context limits not reconciled: the Talker's long-context stage extends to 64k tokens (§4.2) while the Thinker reaches 262,144 (§3)
Pretraining corpus is described for the Omni model only (≈4T tokens, S2); the text model's own data mixture is never disclosed
progressive pre-training curriculum: sequence length raised from 32,768 (S2 general stage) to 262,144 (S3 long-context stage); training occurs at the deployed length, so no post-hoc extension method is described[12]
omnimodal corpus: image-text, video-text, audio-text, video-audio, video-audio-text and pure-text corpora; more than 100 million hours of audio-visual data; AuT audio encoder trained on 40M hours of audio-text pair data generated by Qwen3-ASR; natural-language prompt diversity following Qwen3-Omni; text covers 201 languages/dialects (complete list deferred to the Qwen3.5 report), speech input 113 varieties, speech output 36 varieties[15]
Curriculum / staging
three-stage pre-training: S1 encoder alignment (LLM locked and initialized from Qwen3.5, vision encoder from Qwen3.5, AuT audio encoder; adapters trained before encoders); S2 general stage, all parameters unfrozen, sequence length 32,768; S3 long-context stage, sequence length 262,144 with a higher proportion of long audio/video data[12]
specialist teacher models for agentic, coding and foundational reasoning tasks trained with SFT + RL and distilled into a single unified model; on-policy distillation (OPD) aligns audio-conditioned responses to higher-quality text-conditioned targets; interaction-aligned RL optimizes multi-turn interaction quality (language code-switching, persona consistency, instruction following)[18]
Distillation
specialist distillation (domain-specialized teachers — agentic, coding, reasoning plus vision and audio — distilled into one unified model) and on-policy distillation (OPD) from text-conditioned to audio-conditioned responses; teachers fine-tuned from the pre-trained Qwen-3.5 base checkpoint[18]
hybrid-attention cache: the Gated Delta Net (GDN) module is credited with reducing KV-cache I/O overhead in long-context inference (§2.5); per-layer cache structure and sizes not disclosed
Quantisation shipped
none disclosed
Speculative decoding
MTP (multi-token prediction) module — a Dense Transformer (Table 1) that predicts residual RVQ codebook tokens frame-by-frame for the Talker's streaming speech generation; no speculative decoding described for the text/Thinker path[21]
Serving optimisations
chunked pre-filling of audio/vision encoder outputs; streaming Thinker input and streaming Talker; internal vLLM with torch.compile and CUDA Graph acceleration for the MTP module and codec decoder; first-packet latency Plus 435 ms (audio) / 651 ms (video), Flash 235 ms / 426 ms; Thinker TPS 57–266 (Plus) and 177–942 (Flash) across 1–8 concurrency[22]
Evaluation
Benchmarks (report's own numbers only)
Benchmark
Value
Ref
MMLU-Pro
86.8
Table 4 (Qwen3.5-Plus-Instruct column)
MMLU-Redux
94.3
Table 4 (Qwen3.5-Plus-Instruct column)
SuperGPQA
67.4
Table 4 (Qwen3.5-Plus-Instruct column)
C-Eval
92.3
Table 4 (Qwen3.5-Plus-Instruct column)
IFEval
89.7
Table 4 (Qwen3.5-Plus-Instruct column)
IFBench
51.1
Table 4 (Qwen3.5-Plus-Instruct column)
AA-LCR
62.0
Table 4 (Qwen3.5-Plus-Instruct column)
LongBench v2
60.2
Table 4 (Qwen3.5-Plus-Instruct column)
GPQA
85.9
Table 4 (Qwen3.5-Plus-Instruct column)
LiveCodeBench v6
67.1
Table 4 (Qwen3.5-Plus-Instruct column)
HMMT Nov 25
86.2
Table 4 (Qwen3.5-Plus-Instruct column)
IMOAnswerBench
68.3
Table 4 (Qwen3.5-Plus-Instruct column)
BFCL-V4
66.1
Table 4 (Qwen3.5-Plus-Instruct column)
TAU2Bench
82.7
Table 4 (Qwen3.5-Plus-Instruct column)
Provenance references
stated · §2.3 (Qwen3.5 tokenizer); §2.5 (Hybrid MoE architecture introduced in Qwen3.5); §3 (S1: LLM initialized with parameters from Qwen3.5); §5.1.1 (Qwen3.5-Plus-Instruct)