Qwen3.5

2026-02-16 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba launches Qwen3.5, a natively multimodal generation whose open-weights flagship Qwen3.5-397B-A17B (17 billion active parameters) is released under Apache 2.0, followed by a range from 0.8B to 122B.

### Context Qwen3 (2025) still separated text and vision models (Qwen3-VL). Against GPT-5.2, Claude 4.5 Opus and Gemini 3 Pro, the best closed models kept a lead on most agentic tests. Qwen3.5 aims to bring open models closer to closed ones with a single multimodal model, and is released at the same time as Qwen3.5-Plus, the hosted version (proprietary on the service side) corresponding to the same model. ### What the model brings - Unified vision-language foundation: early-fusion training on multimodal tokens; the vendor claims parity with Qwen3 in text and better results than Qwen3-VL in vision. - Hybrid architecture: Gated DeltaNet (linear attention) combined with a sparse MoE, for high throughput and reduced latency. - Language coverage extended to 201 languages and dialects (vendor announcement). - Full family within a few days: 397B-A17B (February 16), then 122B-A10B, 35B-A3B and 27B (February 24), then 9B, 4B, 2B and 0.8B (February 26 to 28), Hugging Face publication dates. - Hosted version Qwen3.5-Plus: 1M-token context by default, built-in tools, adaptive tool use. ### Architecture and specifications - Qwen3.5-397B-A17B: 397B in total, 17B active; 60 layers; layout 15 x (3 x (Gated DeltaNet then MoE) then 1 x (Gated Attention then MoE)); 512 experts, 10 routed plus 1 shared. - Context: 262,144 tokens natively, extensible up to 1,010,000. - Modalities: text, image, video input; text output. - License: Apache 2.0, with no active-user clause. - Weights: Hugging Face, Qwen/Qwen3.5-397B-A17B (FP8 and GPTQ-Int4 variants published). Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Qwen3.5-Plus, official API, input (USD / M tokens) | 0.4 | | Qwen3.5-Plus, official API, output (USD / M tokens) | 2.4 | | Explicit cache (creation / read) | 0.5 / 0.04 | | Self-hosting | Not disclosed; FP8 and GPTQ-Int4 weights exist to reduce memory | Pricing recorded from the Qwen Cloud page consulted on 2026-09-29, not necessarily the launch pricing. ### Benchmarks | Test | GPT5.2 | Claude 4.5 Opus | Gemini-3 Pro | Qwen3.5-397B-A17B | |---|---|---|---|---| | GPQA | 92.4 | 87.0 | 91.9 | 88.4 | | MMLU-Pro | 87.4 | 89.5 | 89.8 | 87.8 | | IFBench | 75.4 | 58.0 | 70.4 | 76.5 | | SWE-bench Verified | 80.0 | 80.9 | 76.2 | 76.4 | | Terminal Bench 2 | 54.0 | 59.3 | 54.2 | 52.5 | | LiveCodeBench v6 | 87.7 | 84.8 | 90.7 | 83.6 | | BFCL-V4 | 63.1 | 77.5 | 72.5 | 72.9 | Source: Hugging Face model card of Qwen3.5-397B-A17B, scores measured and published by the vendor. ### Versus closed models On this table, Qwen3.5 is ahead of the three cited closed models on IFBench (76.5 versus 75.4 for GPT5.2) and ahead of Claude 4.5 Opus and Gemini 3 Pro on GPQA without surpassing GPT5.2. It remains behind on SWE-bench Verified (76.4 versus 80.9 for Claude 4.5 Opus) and on Terminal Bench 2 (52.5 versus 59.3). The agentic coding gap therefore remains a few points against the best closed model of the period, Claude 4.5 Opus being, in this table, the strongest on code. On price, the hosted version Qwen3.5-Plus is billed at 0.4 USD input and 2.4 USD output per million tokens (page consulted in September 2026), and the vendor also releases the weights under Apache 2.0. For an open-versus-closed choice, this offers a self-hostable option with no royalties, with a quantified trade-off of a few points on agentic coding. ### Key takeaways Qwen3.5 marks Qwen's shift to a single multimodal, hybrid model, with a full range under Apache 2.0. It places open models within a few points of the closed models of February 2026 on most published tests.

Tags : LLM, Open Source, Alibaba, Qwen, Apache 2.0, MoE, Multimodal, Vision