Sulat.com
AI models
OVHcloud AI Endpoints logo

Model details

Qwen3.5-397B-A17B

Qwen3.5-397B-A17B is the first open-weight release in the Qwen3.5 series and is positioned by its creators as a native vision-language model rather than a text-only system retrofitted with vision adapters. It pairs text and image inputs with text outputs, and is designed to perform across a broad task surface that includes reasoning, coding, agent workflows, image understanding, video understanding, and graphical user interface interactions. The expanded language coverage, extending from 119 to 201 languages and dialects, broadens its applicability for global teams that need a single model to serve multilingual and multimodal workloads without stitching together separate specialists.

Under the hood, the model uses a hybrid architecture that fuses a linear attention mechanism implemented through Gated Delta Networks with a sparse mixture-of-experts design. This combination keeps the total parameter count high while activating only a fraction of those parameters per forward pass, which is the key to its favorable efficiency profile relative to dense models of similar capability. The result is a system that aims for strong generalization on agent-style tasks and tool-driven code generation, making it a practical choice for developer-facing assistants, automation pipelines, and enterprise applications that benefit from open-weight deployment alongside hosted access.

OVHcloud AI Endpointsqwen3.5-397b-a17b

Quick Info

Powered by
Provider
OVHcloud AI Endpoints
Model key
qwen3.5-397b-a17b
Release date
May 18, 2026
Last updated
May 18, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost
$0.71
Output token cost
$4.25

Limits

Output tokens
262,144 tokens
Context window
262,144 tokens

Latest news about Qwen3.5-397B-A17B

OVHcloud AI Endpoints

CoverageBenchmark

SemiAnalysis's InferenceX published a technical breakdown of Qwen3.5-397B-A17B covering architecture, evaluations, and inference performance. The article confirms the model has 397B total parameters with 17B activated per forward pass, is distributed under Apache 2.0 on Hugging Face, and is compatible with Transformers The piece notes that Qwen3.5-397B-A17B operates in thinking mode by default, emitting think-tagged content before answers, and unlike Qwen3 drops the /think and /nothink soft switches—non-thinking responses are obtained via API parameters. Early-fusion multimodal training achieves cross-generational parity with Qwen3 w

OVHcloud AI Endpoints

Coverage

Alibaba officially announced the open-source release of Qwen3.5 on February 16, 2026, starting with Qwen3.5-397B-A17B (also branded "Qwen3.5-Plus"). The company described it as a natively multimodal foundation model trained on trillions of vision-language tokens spanning multilingual text, images, videos, STEM, and rea The launch frames Qwen3.5 as a step toward agentic AI that is efficient enough to deploy at scale, with Alibaba emphasizing that multimodal models are moving from demos to production and that inference cost is now the limiting factor. The press release positions Qwen3.5-397B-A17B as rivaling leading frontier models in

Videos about Qwen3.5-397B-A17B