AI Briefing
KO

35B-class performance from 3B

·2026.04.16 23:04

Key point

Qwen released a MoE model with 35B total parameters and 3B active parameters under Apache 2.0.

Details

Released Qwen3.6-35B-A3B as open source.

  • Uses a Sparse MoE architecture with 35B total and 3B active parameters.
  • Licensed under Apache 2.0.
  • Emphasizes agentic coding performance, aiming to match models with 10x more active parameters.
  • Supports multimodal perception and reasoning capability, offering both thinking / non-thinking modes.

Release links include Blog, Qwen Studio, HuggingFace, and ModelScope.

open-sourcemultimodalqwenmoeagentic-codinglarge-language-model
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditSenseNova-U1.5-8B-MoT Released: Unified Text-to-Image and Understanding Model Initialized from Qwen3RedditMica v0.1 4B Released: Open-Source Decision Model for Agent Loops Trained for Under $30GeekNewsXiaomi Releases Open-Weight Multimodal Model MiMo-V2.6-Pro… Higher Intelligence and Competitive Pricing Compared to Similar-Sized ModelsVercel BlogXiaomi Releases Three MiMo V2.6 Models on AI GatewayPyTorchKRTrillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative DecodingHacker NewsApple Releases LensVLM-9B for Processing Compressed Text Images