Qwen3.5-35B-A3B — Model Documentation

Published by Nextbit256 S.L. | Last updated: May 2026 Compliance reference: AI Act Regulation (EU) 2024/1689, Art. 53.1.c and 53.1.d

Model Identification

Field	Detail
Full model name	Qwen3.5-35B-A3B
Developed by	Qwen Team, Alibaba Cloud
Release date	February 2026
Model card	huggingface.co/Qwen/Qwen3.5-35B-A3B
Technical report	Not published as of May 2026 (blog post available: qwen.ai/blog?id=qwen3.5)
Developer blog	qwen.ai/blog?id=qwen3.5

Architecture and Parameters

Field	Detail
Architecture	Mixture of Experts (MoE) with hybrid attention (GatedDeltaNet + Gated Attention)
Total parameters	35B
Active parameters per token	3B (8 routed experts + 1 shared expert activated, out of 256 total experts)
Non-embedding parameters	Not published by the Qwen Team
Layers	40 (layout: 10 × [3 × (GatedDeltaNet → MoE) + 1 × (Gated Attention → MoE)])
Attention	Hybrid: GatedDeltaNet (linear attention) + Gated Attention (quadratic); GQA: 16 Q heads, 2 KV heads
Native context length	262,144 tokens
Extended context (YaRN)	Up to 1,010,000 tokens
Precision	BF16
Modalities	Text, images, video (multimodal)

Apache 2.0 is a permissive open-source license that allows free commercial use, distribution, and modification, provided that the original copyright notice and license text are retained and any modifications are documented.

Nextbit's verification: Nextbit has reviewed the license terms applicable to Qwen3.5-35B-A3B and confirmed that serving this model via API under a commercial inference service is permitted under Apache 2.0. No separate commercial agreement with Alibaba Cloud / Qwen Team is required for this use case.

Restrictions relevant to users: None specific to this model beyond standard Apache 2.0 terms. Users of Nextbit's API who integrate model outputs into their own products remain responsible for their own compliance with applicable law and any downstream licensing requirements.

Training Data Summary

The following is based on publicly available information published by the Qwen Team regarding training data for the Qwen3.5 model family.

Pretraining data

The Qwen3.5 model family was pretrained on a large multilingual corpus covering 201 languages and dialects. The Qwen Team has described the following data categories:

Web content: Text extracted from publicly available web pages, processed and filtered for quality.
Document data: Structured document content (PDF and other formats) processed using multimodal extraction tools.
Synthetic data: Mathematics and programming data generated synthetically to strengthen reasoning and coding capabilities.
Multimodal data: Images and video content for multimodal capabilities.

The model was trained with Multi-Token Prediction (MTP) and reinforcement learning was scaled across large-scale agent environments.

Token volume: Not published by the Qwen Team for the Qwen3.5 family. The Qwen3 generation (February 2025) used approximately 36 trillion tokens; Qwen3.5 likely used a comparable or larger corpus, but no official figure has been released.

Knowledge cutoff

Not published by the Qwen Team. Based on the February 2026 release date, the training data cutoff is estimated to be no later than late 2025. Users should treat this as an approximation.

What is not publicly available

The Qwen Team has not published:

Exact token count for Qwen3.5 pretraining
Composition percentages of data sources
Specific datasets or corpora used
Data filtering or deduplication methodology in detail
Training compute (FLOPs or GPU-hours)

The primary public reference is the Qwen3.5 blog post. A formal technical report had not been published as of May 2026.

Languages Supported

Qwen3.5-35B-A3B supports 201 languages and dialects, as stated in the model card. This includes major world languages across all regions. Full language list: see the Qwen3.5 model card.

Intended Uses

This model is designed for:

Complex reasoning tasks (mathematics, logic, scientific reasoning)
Code generation and analysis
General-purpose instruction following and dialogue
Multilingual tasks and translation
Agentic workflows with tool and function calling
Multimodal tasks: image and video understanding combined with text
Creative writing

The model supports two operation modes:

Thinking mode: Extended internal reasoning before generating a response (uses <think>...</think> tokens), enabled by default; suitable for complex tasks
Non-thinking (instruct) mode: Direct response generation; suitable for conversational and latency-sensitive applications

No explicitly excluded uses are stated in the model card, beyond general responsible AI principles.

Systemic Risk Assessment

Field	Detail
Training FLOPs	Not published by the Qwen Team
Estimated FLOPs (active params method)	~5 × 10²³ (estimate based on 3B active parameters × ~36T tokens as a reference baseline; formula: 6 × N_active × D)
Exceeds 10²⁵ FLOPs threshold?	No — estimated to be approximately 200× below the systemic risk threshold
Note on MoE architecture	FLOPs for MoE models are calculated based on active parameters per token, not total parameters. The ~3B active parameter figure (not the 35B total) is the relevant input for FLOPs estimation.
AI Office designation	Not designated as a systemic risk model as of May 2026
Art. 55 obligations apply?	No

This estimate is based on publicly available information and uses active parameters as the FLOPs basis, consistent with standard MoE evaluation methodology. The training token count for Qwen3.5 has not been officially published; the estimate uses Qwen3's 36T tokens as a conservative reference. If the Qwen Team publishes official training compute figures, this section will be updated.

This documentation is published by Nextbit256 S.L. in accordance with Article 53(1)(c) and 53(1)(d) of Regulation (EU) 2024/1689 (AI Act). Nextbit256 S.L. serves this model via its inference API but did not develop or train it. All training data and architectural information is sourced from public documentation published by the Qwen Team / Alibaba Cloud.

For questions: [email protected]