Run your coding agent on open models.
Pay per token in the cloud, or stay offline on your Mac with a local model. Both take one pip install.
GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and MiniMax M3.1 Flash. No subscription, no card to try it.
Pick how you run.
From $0.10 input, $0.03 cached, $0.50 output per million tokens
Pricing and setupModels
100+ OptiQ models and 80+ community quants on Hugging Face, mixed precision, built for Apple Silicon, and three in the cloud.
In the cloud
| Model | Input | Cached | Output | Context |
|---|---|---|---|---|
| GLM 5.3 Flash | $0.15 | $0.03 | $0.50 | 1M |
| DeepSeek V4.1 Flash | $0.10 | $0.03 | $0.50 | 1M |
| GLM 5.3 | $1.40 | $0.26 | $4.40 | 1M |
Per million tokens.
On your Mac
| Model | Notes | On disk | Capability |
|---|---|---|---|
| Qwen3.8-27B-OptiQ-4bit | dense, vision, MTP | 20.7 GB | 87.98 |
| gemma-4-31B-it-OptiQ-4bit | dense, vision | 23.5 GB | 82.77 |
| MiniCPM-V-4.7-35B-A3B-OptiQ-4bit-REAP-19B | expert-pruned MoE, vision | 14.6 GB | 81.91 |
On disk includes the vision tower and MTP head; Capability Score from each model card.
Measured on real coding tasks.
Each number is from a run we published, with the method on the linked page.
tasks solved with Boost, against 4 for the local model alone, for $2.80 in total.
The 30-task runof SWE-bench-Lite resolved by OptiQ Code on a 4-bit 4B local model, against 25% for mini-swe-agent on the same model.
Harness comparisonUnder the hood: the CLI.
The same engine runs all three. Use it directly for serving, quantizing and pruning models yourself.
CLI reference- optiq serve --model <repo>
- An OpenAI- and Anthropic-compatible server for any client.
- optiq convert <repo>
- Measure each layer's sensitivity and build a mixed-precision quant.
- optiq prune-experts <repo>
- Remove half of a mixture-of-experts model's experts.
- optiq serve --cloud-only
- The cloud models behind a local endpoint, with no model loaded.