Qwen3.5

Qwen Team · 2026-04-21 · analysed 2026-08-10 · skill v0.4.0

hybrid hybrid text-only undisclosed frontier frontier-generalist

Design overview

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

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

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyQwen3.5[1]
OrganisationQwen Team[2]
Release date2026-04-21[3]
Report URLhttps://arxiv.org/abs/2604.15804[4]
LicenseCC BY 4.0[5]
Open weightsnot disclosed

Variants

Scale

Total parametersnot disclosed
Active parameters / tokennot disclosed
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary size250000[6]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
Qwen3.5-Omni-Plusnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed262,144
Qwen3.5-Omni-Flashnot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed262,144
Qwen3.5-Plus-Instruct (text-only counterpart)not disclosednot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed

Core block

Block typehybrid
Attention varianthybrid[7]
Attention layer patternnot disclosed
Depth mixingsequential-residual
Activationnot disclosed
Expert countnot disclosed
Experts per tokennot disclosed
Shared expertsnot disclosed
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
MethodRoPE
RoPE base frequencynot disclosed
Partial RoPEno
Typenot disclosed
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno

Context

Trained context length262,144[10]
Deployed context length262,144[11]
Extension methodprogressive 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]

Tokenizer

Algorithmbyte-level byte-pair encoding (BPE)[13]
NotesQwen3.5 tokenizer with vocabulary 250k up from 150k; report claims 10–60% encoding/decoding efficiency improvement across most languages[6]

Training

Training tokens4T[14]
Data compositionomnimodal 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 / stagingthree-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]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[16]
Preference optimisationother[17]
Reasoning trainingspecialist 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]
Distillationspecialist 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]

Modality

Typetext-only[19]
Attachmentn/a[20]

Inference efficiency

KV-cache designhybrid-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 shippednone disclosed
Speculative decodingMTP (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 optimisationschunked 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)

BenchmarkValueRef
MMLU-Pro86.8Table 4 (Qwen3.5-Plus-Instruct column)
MMLU-Redux94.3Table 4 (Qwen3.5-Plus-Instruct column)
SuperGPQA67.4Table 4 (Qwen3.5-Plus-Instruct column)
C-Eval92.3Table 4 (Qwen3.5-Plus-Instruct column)
IFEval89.7Table 4 (Qwen3.5-Plus-Instruct column)
IFBench51.1Table 4 (Qwen3.5-Plus-Instruct column)
AA-LCR62.0Table 4 (Qwen3.5-Plus-Instruct column)
LongBench v260.2Table 4 (Qwen3.5-Plus-Instruct column)
GPQA85.9Table 4 (Qwen3.5-Plus-Instruct column)
LiveCodeBench v667.1Table 4 (Qwen3.5-Plus-Instruct column)
HMMT Nov 2586.2Table 4 (Qwen3.5-Plus-Instruct column)
IMOAnswerBench68.3Table 4 (Qwen3.5-Plus-Instruct column)
BFCL-V466.1Table 4 (Qwen3.5-Plus-Instruct column)
TAU2Bench82.7Table 4 (Qwen3.5-Plus-Instruct column)

Provenance references

  1. 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)
  2. stated · Title page byline
  3. stated · arXiv header line: 'arXiv:2604.15804v2 [cs.CL] 21 Apr 2026'
  4. stated · pipeline
  5. stated · arXiv HTML header: 'License: CC BY 4.0'
  6. stated · §2.3
  7. stated · §1 ('Hybrid-Attention Mixture-of-Experts (MoE) framework'); §2.5 (GDN module)
  8. n/a · §3 (no extension method applies; training occurs at the deployed length)
  9. stated · §1 ('256k-token long-context input'); §3 (S3 length 262,144)
  10. stated · §3 (S3: 'we increased the maximum token length from 32,768 to 262,144')
  11. stated · Abstract ('supports a 256k context length'); §1 ('256k-token long-context input')
  12. stated · §3
  13. stated · §2.3 ('adopts byte-level byte-pair encoding')
  14. stated · §3 (S2: 'approximately 4 trillion tokens')
  15. stated · §1, §2.2, §3, Table 3
  16. stated · §4.1 (Stage 1: specialist teachers trained via independent Supervised Fine-Tuning (SFT) and RL; ChatML-format corpus)
  17. stated · §4.1 Stage 3 (Interaction-Aligned RL, algorithm unnamed); §4.2 (Talker: DPO + GSPO with rule-based rewards)
  18. stated · §4.1
  19. stated · §5.1.1 ('Qwen3.5-Plus-Instruct' as the text-only counterpart, Table 4)
  20. n/a · §5.1.1 (record is the text-only model)
  21. stated · §2.4, Table 1
  22. stated · §2.5, Table 2