Qwen Rolls Out New Vision‑Language Model To Advance Coding, Reasoning, And Multimodal AI Performance

by
Alisa Davidson

Revealed: February 16, 2026 at 9:30 am Up to date: February 16, 2026 at 9:04 am

by Ana

Edited and fact-checked:
February 16, 2026 at 9:30 am

To enhance your local-language expertise, typically we make use of an auto-translation plugin. Please notice auto-translation is probably not correct, so learn authentic article for exact data.

Alibaba Cloud’s Qwen crew has launched the primary mannequin in its new Qwen3.5 sequence, unveiling the open‑weight Qwen3.5‑397B‑A17B. 

Positioned as a local imaginative and prescient‑language system, the mannequin delivers robust efficiency throughout reasoning, coding, agent duties, and multimodal understanding, reflecting a major advance within the firm’s massive‑scale AI improvement efforts. 

The mannequin is constructed on a hybrid structure that mixes linear consideration by means of Gated Delta Networks with a sparse combination‑of‑specialists design, enabling excessive effectivity throughout inference. Though the complete system incorporates 397 billion parameters, solely 17 billion are activated for every ahead go, permitting it to keep up excessive functionality whereas lowering computational price. The discharge additionally expands language and dialect protection from 119 to 201, broadening accessibility for customers and builders worldwide.

Qwen3.5 Marks A Main Leap In Reinforcement Studying And Pretraining Effectivity

The Qwen3.5 sequence introduces substantial features over Qwen3, pushed largely by in depth reinforcement studying scaling throughout a variety of environments. Somewhat than optimizing for slim benchmarks, the crew targeted on growing process problem and generalizability, leading to improved agent efficiency throughout evaluations equivalent to BFCL‑V4, VITA‑Bench, DeepPlanning, Device‑Decathlon, and MCP‑Mark. Further outcomes shall be detailed in an upcoming technical report.

Pretraining enhancements span energy, effectivity, and flexibility. Qwen3.5 is skilled on a considerably bigger quantity of visible‑textual content knowledge with strengthened multilingual, STEM, and reasoning content material, enabling it to match the efficiency of earlier trillion‑parameter fashions. Architectural upgrades—together with larger‑sparsity MoE, hybrid consideration, stability refinements, and multi‑token prediction—ship main throughput features, significantly at prolonged context lengths of 32k and 256k tokens. The mannequin’s multimodal capabilities are strengthened by means of early textual content‑imaginative and prescient fusion and expanded datasets masking photos, STEM supplies, and video, whereas a bigger 250k vocabulary improves encoding and decoding effectivity throughout most languages.

The infrastructure behind Qwen3.5 is designed for environment friendly multimodal coaching. A heterogeneous parallelism technique separates imaginative and prescient and language elements to keep away from bottlenecks, whereas sparse activation allows close to‑full throughput even on blended textual content‑picture‑video workloads. A local FP8 pipeline reduces activation reminiscence by roughly half and will increase coaching pace by greater than 10 %, sustaining stability at huge token scales. 

Reinforcement studying is supported by a totally asynchronous framework able to dealing with fashions of all sizes, enhancing {hardware} utilization, load balancing, and fault restoration. Strategies equivalent to FP8 finish‑to‑finish coaching, speculative decoding, rollout router replay, and multi‑flip rollout locking assist keep consistency and cut back gradient staleness. The system is constructed to help massive‑scale agent workflows, enabling seamless multi‑flip interactions and broad generalization throughout environments.

Customers can work together with Qwen3.5 by means of Qwen Chat, which gives Auto, Considering, and Quick modes relying on the duty. The mannequin can be out there by means of Alibaba Cloud’s ModelStudio, the place superior options equivalent to reasoning, internet search, and code execution may be enabled by means of easy parameters. Integration with third‑social gathering coding instruments permits builders to undertake Qwen3.5 into current workflows with minimal friction.

In accordance with the Qwen crew, Qwen3.5 establishes a basis for common digital brokers by means of its hybrid structure and native multimodal reasoning. Future improvement will deal with system‑stage integration, together with persistent reminiscence for cross‑session studying, embodied interfaces for actual‑world interplay, self‑directed enchancment mechanisms, and financial consciousness for lengthy‑time period autonomous operation. The target is to maneuver past process‑particular assistants towards coherent, persistent brokers able to managing advanced, multi‑day targets with dependable, human‑aligned judgment.

Alisa, a devoted journalist on the MPost, focuses on cryptocurrency, zero-knowledge proofs, investments, and the expansive realm of Web3. With a eager eye for rising traits and applied sciences, she delivers complete protection to tell and have interaction readers within the ever-evolving panorama of digital finance.

Extra articles