Open models for coding agents, paid per token.
GLM 5.3 and DeepSeek V4 Pro for the hardest tasks, MiniMax M3 in between, the Flash models and Qwen3.8 27B for fast, cheap ones, and MiniMax M3.1 Flash when the task includes images or video. No subscription.
Use them from the agent you already use, from OptiQ Code with Boost, or from the API. No card needed to start.
https://cloud.mlx-optiq.com/api/v1
Eight models, chosen and tested for coding agents
Prices are per million tokens, and output includes reasoning. A long agent run resends its conversation on every call and most of it is served from cache, so the cached price matters most.
| Model | Id | Input | Cached input | Output | Context | Typical speed |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | deepseek-v4.1-flash | $0.10 | $0.03 | $0.50 | 1M | — |
| GLM 5.3 Flash | glm-5.3-flash | $0.15 | $0.03 | $0.50 | 1M | — |
| MiniMax M3.1 Flash | minimax-m3.1-flash | $0.15 | $0.02 | $0.60 | 1M | — |
| Qwen3.8 27B | qwen3.8-27b | $0.15 | $0.03 | $2.00 | 256K | — |
| MiMo V2.6 Flash | mimo-v2.6-flash | $0.22 | $0.01 | $0.65 | 256K | — |
| MiniMax M3 | minimax-m3 | $0.50 | $0.10 | $2.00 | 512K | — |
| GLM 5.3 | glm-5.3 | $1.40 | $0.26 | $4.40 | 1M | — |
| DeepSeek V4 Pro | deepseek-v4-pro | $2.20 | $0.08 | $6.60 | 1M | — |
All OptiQ Cloud models are served at full precision, and your requests are never used for training. GLM 5.3 and DeepSeek V4 Pro are the larger models, for the hardest tasks. MiniMax M3 sits between them and the fast, cheap ones: the four Flash models and Qwen3.8 27B, the same model as the OptiQ quant you can run on your Mac. Boost runs on GLM 5.3 Flash unless you pick another model. How the models are served.
Use it from your agent
Sign in once per Mac, then one command starts an agent on a cloud model.
optiq run starts optiq serve --cloud-only, which loads no
local model, and configures the agent for that session only. The agent talks to
optiq serve, which holds the key, so the agent never sees it.
$ pip install -U mlx-optiq
$ optiq cloud login
$ optiq run claude --cloud glm-5.3-flashClaude Code
optiq run claude --cloud glm-5.3-flash
Codex
optiq run codex --cloud glm-5.3-flash
OpenCode
optiq run opencode --cloud glm-5.3-flash
Mistral Vibe
optiq run vibe --cloud glm-5.3-flash
Hermes Agent
optiq run hermes --cloud glm-5.3-flash
OpenClaw
optiq run openclaw --cloud glm-5.3-flash
Any client pointed at optiq serve can
also pick a cloud model by name, such as optiq-cloud/glm-5.3-flash or
optiq-cloud/glm-5.3, and the OptiQ Lab lists them in
its model picker. Setup for each client.
Hard tasks in OptiQ Code: the cloud leads, your Mac helps
Type /boost before a hard task. From the next task the cloud
model leads and your local model is its local tool: it reads files and
answers questions, runs the tests and reports the failures, and makes edits the cloud
model has specified. Only the short answer goes back to the cloud.
$ optiq code
/boost
add retry with backoff to the HTTP client and make the tests pass/boost glm-5.3 picks the
model and /boost off stops it. If the cloud cannot answer, the task
carries on with your local model. The run summary shows what the task cost and what is
left on your balance.
| 30 real tasks | Local model alone | With Boost |
|---|---|---|
| Tasks solved | 4 / 30 | 15 / 30 |
| DeepSWE-mini target tests | 313 / 771 | 706 / 771 |
| Terminal-Bench tasks solved | 4 / 14 | 8 / 14 |
| Cost for all 30 tasks | free, local | $2.80 |
The same 30 tasks from DeepSWE-mini and Terminal-Bench mini, the same local Qwen3.6-35B-A3B on a 24 GB Mac mini, run in OptiQ Code with and without Boost, one attempt per task. The full write-up is on the blog.
Ask on an issue, get a pull request back
Install the optiq-bot app, add your Cloud key as a repository secret, and mention @optiq-bot on an issue. OptiQ Code works on a branch and opens a pull request. It carries on when someone comments, reviews it, or CI fails, and a person still decides what gets merged. Runs are billed per token from your balance. Set it up.
Or call the API directly
OpenAI's chat completions format, with streaming. Create a key on the dashboard; it is shown once. The API reference has the rest.
https://cloud.mlx-optiq.com/api/v1
from openai import OpenAI
client = OpenAI(base_url="https://cloud.mlx-optiq.com/api/v1", api_key=OPTIQ_CLOUD_API_KEY)
reply = client.chat.completions.create(
model="deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Hello"}],
)Standard settings are accepted, from
temperature to reasoning_effort; settings you leave out use the
ones the model was tested with. Each reply's usage says what the call cost
and the balance left. API reference.
Limits you set, and no subscription
Each API key has a spending limit and rate limits you can set in the dashboard. You pay for the tokens you use, from a balance you top up when you want.
| Per key | Default | You can |
|---|---|---|
| Spending | $100 a month | raise it, lower it or remove it |
| Requests | 60 a minute | change it |
| Calls at once | 10 | change it |
| Models | all | narrow it to some |
| Expiry | none | set a date |
$2 to start
No card needed.
Top up any amount
From $10. Credit lasts 12 months.
What leaves your machine
What leaves my machine?
The conversation of a task on a cloud model: your request, the replies, and the tool results it asks for, including the parts of files it reads. Tasks on your local model never reach the network, and neither do the jobs Boost hands to your local model, only their short answers. Your repository is not uploaded, indexed or scanned.
Is my code used for training?
No. OptiQ Cloud does not store prompts or replies, and your requests are never used for training. What is kept is token counts, timings and cost, so the dashboard can show what a task cost.
What if my balance runs out?
The call is refused with a link to top up. Nothing is charged past an empty balance except the call already in progress, and in OptiQ Code your local model keeps working without the cloud.
Can I turn it off?
Nothing reaches the cloud until you pick a cloud model or Boost.
optiq cloud logout revokes the key rather than only forgetting it.
Run your next task on an open model
$2 of credit to start, and one command for whichever agent you already use.