Qwen 3.8 27B (24K, no thinking)
The free one. Fits entirely on a 16 GB card at 24K context with thinking turned off.
Tested 2026-11-08
This is a 27B model running locally through Ollama, under the tag qwen38-24k, at 24K context with thinking disabled. The point of the setup is that the whole model and the whole context window fit on the GPU: a GeForce RTX 4080 SUPER with 16 GB, in a machine with 60 GB of system RAM.
Nothing leaves the machine and nothing costs money per token. It is slower than a cloud model and it is worse at long jobs. It is also the model I use most often, because most of my tasks are small.
Pros
- No per-token cost and no account, ever
- Works offline, including on a plane
- Fast enough for single-file fixes and explanations
- 24K of context that does not spill into system RAM
Cons
- Noticeably weaker on multi-step reasoning than the paid models
- 24K context fills up on real projects and then it starts forgetting
- Thinking off means less care on hard questions
What I tested
| Test | Result |
|---|---|
| Long document summary | TODO: run the test doc and record where it starts dropping detail. |
| Coding task | Good on single files and small scripts. Fine on this static site's build script when asked for one change at a time. |
| Hard reasoning | TODO: record the question and result. |
| Image understanding | TODO: not tested yet. |
| Cost for a week of real use | $0.00. It runs on the card already in the machine. The real cost is electricity and the wait. |
Bottom line: If the task fits in one file, start here. It costs nothing and it is fast. Move up only when it visibly struggles.
What this site is
I test AI tools myself — cloud models, local models, and the apps that manage them. Reviews only appear once I’ve run the same prompts through each tool on my own machine or account. No sponsorships, no affiliate links, no ads. More about how this works.
One email a week
What I tested this week, what it cost to run, and the one thing worth your time. No spam, unsubscribe in one click.