Qwen3.6-35B-A3B is an open-weight model from Alibaba's Qwen team, released in April 2026 under Apache-2.0. It is a mixture-of-experts model with 35 billion parameters, 3 billion active per token, a 262K context and image input. The 4-bit Ollama build is 23 GB, so plan a 32 GB GPU or a large-memory Mac.
Other good fits
Reading long documents, screenshots and scanned pages with one model: 262,144 tokens of context plus image input.
+ Pros
Read pros
Pick it when you want a capable model on your own hardware for coding, drafting and analysis, under a licence that allows commercial use and changes.
Pick it for speed. Only about 3 billion of its 35 billion parameters work on each token, so it answers quickly for its size.
Pick it when you need long input or pictures: 262,144 tokens natively (extendable to about 1 million) and a vision encoder for screenshots, charts and scans.
Pick it when you want a choice of runtime. The official weights work with vLLM, SGLang and Transformers; ready-made builds exist for Ollama, llama.cpp (GGUF) and Apple's MLX.
− Cons
Pick something else when coding agents are the main job: Ornith-1.5-35B-A3B, a model of the same size and design tuned for agent work, scores higher on coding benchmarks in its publisher's tests.
Pick something else when memory is tight: Phi-4-mini is 2.5 GB in Ollama and runs on a laptop CPU.
Pick something else when you'd rather not run hardware at all: OpenRouter gives you hosted open models through one API.
Avoid if
You want the newest Qwen generation. Qwen3.8-27B (August 2026, Apache-2.0) is newer, though it is a dense model and slower per token.
Context
Documented gap:qwen3-6 is a legacy catalogue ID, not an exact official model identifier confirmed in the checked Qwen sources. It must not be silently equated with Qwen3, Qwen3.5 or any size/quantization variant.
Requirements
Before use, record the publisher model ID, revision, licence, runtime, precision, context and hardware target. The checked sources do not establish these facts for this legacy ID.
Licence & cost
Prices checked 23 September 2026: free weights (Apache-2.0); you pay for the machine.
Ollama's default 4-bit build is 23 GB. Plan a GPU with 32 GB, or a Mac with 48 GB of unified memory.
The official 16-bit weights need about 70 GB; Qwen's FP8 version roughly halves that, for a server with one 48-80 GB GPU.
Long contexts need extra memory. Start at 32K tokens and raise it only when a task needs it.
Running effort is low on Ollama: pull an updated build now and then. On vLLM or SGLang, use recent versions, because its hybrid attention design needs current support.
First task
Documentation-based suggestion — not run here.
Turn 20 scanned supplier invoices into a spreadsheet, without uploading them anywhere.
Install Ollama and run ollama pull qwen3.6:35b (23 GB).
For each invoice image, run: ollama run qwen3.6:35b "Extract supplier, invoice number, date, net amount, VAT and total as one CSV line. Write UNKNOWN for anything you can't read. ./invoice-01.png"
Paste the lines into a spreadsheet.
Check five invoices against the paper copies before you trust the rest. If a field is often wrong, describe where it sits on the page in the prompt and run it again.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Based on Qwen's model card and the model list on Hugging Face, the Qwen3.6 blog post and Ollama's qwen3.6 page, checked 23 September 2026. We did not run it for this review. The catalogue ID qwen3-6 maps to Qwen/Qwen3.6-35B-A3B.
Qwen3.6-35B-A3B model card · Sep 23, 2026 — Apache-2.0; 35B total and 3B activated parameters; vision encoder; 262,144-token context, extensible to about 1,010,000; compatible with Transformers, vLLM, SGLang; first published 15 April 2026.
Qwen3.6-35B-A3B blog post · Sep 23, 2026 — Qwen's release announcement for the model.
Ollama: qwen3.6 · Sep 23, 2026 — qwen3.6:35b build is 23 GB with a 256K context and text and image input.
Run an open mixture-of-experts model with coding and multilingual capabilities.
Use its long context window for substantial documents or codebases.
Apple MLX support makes it a practical local option on suitable Macs.