A free model you can download right now runs on your gaming PC and matches Claude Sonnet 4.5 on benchmarks. No API bill. No subscription. Just download and go.
Alibaba released the Qwen3.5-Medium series on February 24, and the flagship 35B-A3B model is turning heads in the local AI community. The headline claim: frontier-class performance that fits in 32GB of VRAM.
The Numbers
The Qwen3.5-35B-A3B uses a sparse Mixture-of-Experts architecture. It has 35 billion total parameters but only activates 3 billion per token - that’s where the efficiency comes from.
On benchmarks, it beats or matches Claude Sonnet 4.5 across multiple categories:
- MMLU-Pro (knowledge): 85.3%
- MMMU-Pro (visual reasoning): 75.1%
- SWE-bench Verified (real-world coding): 69.2%
- LiveCodeBench v6 (competitive programming): 74.6%
The model particularly shines on visual reasoning and document understanding - categories where it trades blows with GPT-5 mini and comes out ahead in some tests.
Context length stretches to 262K tokens natively, with YaRN scaling extending that to over 1 million tokens if you have the VRAM.
Running It Locally
Here’s what the hardware requirements actually look like:
RTX 4090 (24GB): With 4-bit quantization, you can run the full model with reduced context. Users report 72.9 tokens per second - fast enough for interactive chat.
RTX 5090 or 32GB cards: Full 1-million-token context with 4-bit quantization. This is the sweet spot.
Apple Silicon (M3/M4 Max): Over 100 tokens per second on M4 Max with 4-bit quantization using MLX optimization. The unified memory architecture helps significantly.
The 4-bit quantization is what makes this work. Alibaba claims it’s “near-lossless” - meaning compressed performance almost matches the full-precision version.
Setup is straightforward:
ollama pull qwen3.5:35b-a3b
ollama run qwen3.5:35b-a3bOr through LM Studio, llama.cpp, or the SGLang framework for more control.
The Full Lineup
The Medium series includes four models:
| Model | Total Params | Active Params | Use Case |
|---|---|---|---|
| Qwen3.5-Flash | Hosted | - | API access, 1M context |
| Qwen3.5-35B-A3B | 35B | 3B | Best efficiency/quality ratio |
| Qwen3.5-122B-A10B | 122B | 10B | Higher quality, more VRAM |
| Qwen3.5-27B | 27B | 27B | Dense model, maximum quality per param |
All ship under Apache 2.0. Download, modify, deploy commercially - no restrictions.
What It’s Good At
The 35B-A3B excels at:
- Coding: 69.2% on SWE-bench Verified is competitive with models that cost money per token
- Document analysis: Native multimodal support handles images and video
- Long-context work: 262K native context, extensible to 1M+
- Tool calling: Built-in agentic capabilities
The model supports 201 languages and dialects, and includes both “thinking mode” (extended reasoning) and “instruct mode” (direct responses).
The Caveats
We covered the censorship patterns in Qwen models when Qwen 3.5 launched earlier this month. The short version: instruct-tuned versions contain documented alignment to Chinese government positions on topics like Taiwan, Tiananmen Square, and Xinjiang.
For coding, math, and non-political tasks, you’ll likely never encounter these limitations. For journalism, research, or applications needing geopolitically neutral responses, consider using base (non-instruct) models or community fine-tunes.
Why This Matters
Six months ago, running a model competitive with Claude Sonnet required either paying per token or owning datacenter hardware. Now you can download one for free and run it on a gaming PC.
The cost differential is stark. Qwen3.5-Flash delivers frontier-adjacent performance at $0.10 per million input tokens - roughly 1/13th the cost of Claude Sonnet 4.6. The local models cost nothing after download.
This doesn’t mean paid APIs are obsolete. Anthropic’s Claude and OpenAI’s GPT-5 still lead on some benchmarks, particularly complex reasoning tasks. And local inference requires hardware, electricity, and technical setup.
But the gap is narrowing. Fast.
The Bottom Line
Qwen3.5-35B-A3B delivers frontier-class AI performance that runs on consumer hardware. The model is free, open-source, and available right now on Hugging Face and Ollama.
If you have an RTX 4090 or better - or an M3/M4 Mac with substantial RAM - you can run AI that benchmarks alongside Claude Sonnet 4.5 entirely locally, with no ongoing costs.
The main trade-off is documented censorship on certain political topics. For most coding, analysis, and general tasks, that won’t matter. Test it yourself and decide.