Guides · dev · run qwen locally
Run Qwen locally when the API price is not the constraint
Local models are a privacy decision that sometimes saves money. The money part is easy to fake with a benchmark screenshot from one person’s GPU. We will not do that. The API price is in the repo. The hardware speed is not, until we measure it.
The price you are replacing
| Model | Blended | vs Opus | vs Sol |
|---|---|---|---|
| qwen3.8-27b | $0.730 | 91% cheaper | 82% cheaper |
| qwen3.8-flash | $0.230 | 97% cheaper | 94% cheaper |
| deepseek-v4.1-flash | $0.198 | 98% cheaper | 95% cheaper |
Qwen3.8 27B is the id in the spotlight that people mean when they talk about a model that might fit a workstation. Flash is so cheap on the API that electricity and your time can cost more than the tokens, unless the prompt cannot leave the building. That is the whole trade.
| Model | Session | vs Opus | vs Sol | Cache price |
|---|---|---|---|---|
| qwen3.8-27b | $0.0088 | 90% cheaper | 80% cheaper | published |
| qwen3.8-flash | $0.0031 | 97% cheaper | 93% cheaper | published |
| Anthropic: Claude Opus 5.5 | $0.088 | — | — | published |
| OpenAI: GPT-6.1 Sol | $0.044 | — | — | published |
The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.
The document-question mix is the fairer local comparison: a pile of context, a short answer. A coding-agent mix of 3 million tokens is not what a laptop context window is for.
How to do it without a made-up speed
Qwen’s Hugging Face org is Qwen. DeepSeek’s is deepseek-ai. Kimi K3’s card describes a 2.8T model. That is not this guide. The licence note is where K3 belongs.
What X threads get wrong
A post that says a 27B model “flies” on a specific Mac is one machine, one quant, one prompt length. It is not a spec. If we later measure a machine we control, the number will be dated and the prompt will be attached. Until then the chart above is the only figure.
Related: hosts outside China, Qwen API.
Questions
Can I run the flagship on a laptop?
Not the largest Qwen or Kimi K3. Start from a small instruct id whose weights you can actually download. We are not publishing an unverified tokens-per-second claim for a 16 GB Mac.
When is local cheaper?
When the machine is already paid for and the prompts are sensitive. Otherwise the flash API price in the chart is the number to beat.
Sources
- OpenRouter models API, accessed 2026-10-16
- Qwen on Hugging Face, accessed 2026-10-16
- DeepSeek on Hugging Face, accessed 2026-10-16