Model · Alibaba · Text
Qwen3.6 35B-A3B
Alibaba's mid-size open-weight mixture-of-experts from the Qwen3.6 generation — 35B total / 3B active parameters, Apache 2.0, natively multimodal (text + vision) with up to ~1M-token context. Still the current Qwen MoE at this size: Qwen3.8 has no MoE this small.
- Modality
- Text
- License
- Apache 2.0 (Open)
- Parameter size
- 35B (3B active)
- Context window
- 262,144 tokens · up to 1.01M 262,144 tokens natively, extensible to about 1,010,000 with YaRN scaling, per the model card.
- Released
- April 17, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- vision
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 32 GB (48 GB recommended) of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 24 GB (32 GB recommended) of video memory.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
All 35 billion parameters must sit in memory even though only 3 billion work at a time, so you need a 32 GB Mac; it then runs quickly.
Numbers are for: 4-bit build of the 35B-A3B (about 21 GB). Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Apache 2.0 — no commercial restrictions
- Strong multilingual and Chinese-language performance
- MoE architecture — fast inference for the parameter count
- Native vision capability
- Strong coding and math performance
Weaknesses
- Needs serious GPU footprint for self-hosting
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | Apache 2.0; FP8 and quantized variants available |
| Ollama | local | Quantized variants available |
Official sources
- Model card model
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: estimate from size; mlx-community/Qwen3.6-35B-A3B-4bit; Ollama qwen3.6 page.
- — Un-deprecated and licence corrected. The 2026-09-08 supersession by Qwen3.8 27B was wrong: Qwen3.8 ships a 27B dense model, a 125B-A6B Flash-Next MoE and a 2.4T MoE, so nothing replaces a 35B-A3B MoE, and Qwen3.6-35B-A3B is still heavily used (4.21M Hugging Face downloads/month, 3,044 likes). The licence was also wrong: the Hugging Face model card and LICENSE file are Apache 2.0, not a Qwen License with a 100M-MAU cap, so license_type is now open. Context window (262,144 native, ~1.01M with YaRN) added from the card.
- — Superseded by Qwen3.8 27B (August 2026). Note the filename is a legacy artifact of the corrected 2026-06-08 entry and no longer matches the title.
- — Replaced fabricated 'Qwen 3.5 122B' entry (no such repo) with the real current open-weight flagship Qwen3.6-35B-A3B (released 2026-04-17). The real Qwen3.5 flagship was Qwen3.5-397B-A17B (2026-02-16).
- — Initial entry. Sighted on Hugging Face (lastModified 2026-04-24).
Esc