mlx-optiq
Benchmark · October 15, 2026

Boost took a Mac mini from 4 to 15 solved coding tasks, for $2.80

A 35B local model on a 24 GB Mac mini solved 4 of 30 real coding tasks on its own, running in OptiQ Code. With Boost turned on, the same agent on the same Mac solved 15. On the DeepSWE-mini repos, the share of hidden tests passing went from 41% to 92%. All 30 tasks together cost $2.80, which is 19 cents per solved task.

Tasks solved · 30 tasks · same agent, same Mac

What Boost does

Boost is one command. Type /boost in OptiQ Code before a hard task, and optiq-boost, the OptiQ Cloud model, takes the lead. It reads the task, plans the change, and writes the edits that matter. Your local model keeps working as its helper. The cloud model can hand it a self-contained job, like reading a set of files and answering a question, or running the tests and reporting only the failures. Commands, tests, and edits all run on your Mac, and the run summary shows what the task cost.

Boost is not tied to OptiQ Code. The optiq serve endpoint answers the model name optiq-boost by sending the call to OptiQ Cloud, so Claude Code, Codex, OpenCode and other agents can use it by picking that model, with no plugin. Every other model name is still answered by your local model. The OptiQ Cloud docs list the setting for each agent.

The results

The 30 tasks are the 16 in DeepSWE-mini, feature requests on real Go, TypeScript and Python repos graded by hidden tests. The other 14 come from Terminal-Bench mini, system jobs in a container such as building an old C program or recovering a database from its write-ahead log. Same tasks, same limits, same local model on the same Mac, with and without Boost.

OptiQ Code, local model alone vs with Boost
Local aloneWith Boost
DeepSWE-mini tasks solved0 / 167 / 16
DeepSWE-mini target tests313 / 771 (40.6%)706 / 771 (91.6%)
Terminal-Bench tasks solved4 / 148 / 14
Terminal-Bench tests26 / 6052 / 60
All tasks solved4 / 3015 / 30

The seven DeepSWE solves are in koota, textual, yaegi, valibot, aiomonitor, ofetch and narwhals, across TypeScript, Python and Go. On Terminal-Bench, Boost picked up five tasks the local model had missed (qemu-alpine-ssh, build-pmars, mcmc-sampling-stan, rstan-to-pystan and vulnerable-secret) and lost one it had solved, regex-log.

Several of the misses were close. Bandit finished with all 69 of its target tests passing and one existing test broken, and termenv had 32 of 35. A rerun with the cloud model set to think harder did not solve either.

What it cost

Boost bills per token from a prepaid balance, with no subscription: $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. A long agent run re-sends its conversation on every call, so most of the input is cached. The 30 tasks used 60.4 million input tokens, 54.7 million of them from cache, and 0.6 million output tokens. The median task cost $0.056 and the most expensive one $0.33.

At Boost's priceCost
All 30 tasks$2.80
Median task$0.056
Most expensive task (koota, 124 turns)$0.33
Per solved task$0.19

For scale, here is the same token count priced at the listed rates of other models people use for coding, if you called them directly. This prices our workload, not theirs. A different model would take a different number of turns and tokens on the same tasks, and might solve more of them. What it shows is how much the per-token price matters on a long agent run, where nearly all of the input is the conversation being re-read.

The same 61 million tokens, Boost vs other models at their listed rates

Other models at their listed per-token rates on October 15, 2026, applied to the input, cached-input and output tokens of the Boost run.

Why it runs on OptiQ Code

Boost runs inside OptiQ Code, and the agent matters. On the same local model with no cloud, OptiQ Code passed 313 of the 771 hidden DeepSWE-mini tests, against 179 for mini-swe-agent and 42 for opencode. It solved 4 Terminal-Bench tasks against 3 and 2. A small model on a long task loses track of requirements, repeats commands that already failed, and declares the work done with a test still failing. OptiQ Code watches for each of those and steers the model back. That is a good base, but none of the three agents finished a single DeepSWE-mini task with the local model alone. Boost is what gets those tasks done.

Local model only · all 30 tasks
OptiQ Codemini-swe-agentopencode
DeepSWE-mini target tests313 / 771179 / 77142 / 771
Terminal-Bench tasks solved4 / 143 / 142 / 14

How we ran it

Held fixedValue
Local modelQwen3.6-35B-A3B, OptiQ 4-bit
HardwareMac mini M4, 24 GB, served by optiq serve
TasksDeepSWE-mini 16 + Terminal-Bench mini 14
Limits2 hours and 1,000 turns per task, one attempt each
Gradingthe benchmark's own hidden tests

These 30 tasks are the set a cloud model has to clear before we offer it through Boost. Every agent talked to the same optiq serve endpoint, so the local model, its quantization and its KV cache were the same in every run. One attempt per task is a small sample. A single task can swing by many tests from one run to the next. The totals across 30 are steadier than any one of them.

On these tasks the cloud model did most of the work itself, but not by choice. It handed jobs to the local model on all 16 DeepSWE tasks and on 4 of the 14 Terminal-Bench tasks, 25 hand-offs in all. Most of them stopped short: 14 ran out of the helper's turn limit before answering and 8 were cut off at its answer length limit, so the cloud model went back to reading files and running tests itself. Making those hand-offs succeed, which is what saves cloud tokens, is what we are working on next. Earlier versions of Boost let the local model lead and called the cloud in when it stalled. That did worse every time we compared the two, so since 0.6.0 the cloud model leads.

We also counted existing tests each run broke. With Boost, bandit and prometheus each lost one, because the model did not run the full suite before finishing. dynamodb-toolbox lost 953 to a bug of ours: the end-of-run rollback mistook TypeScript type errors for a broken build. 0.6.0 fixes it. Running alone, the local model broke 564 existing tests in expr and 82 in prometheus.

Try it

An OptiQ Cloud account starts with $2 of credit, good for 14 days, with no card needed. That is about 35 tasks at this run's median. After that, credit is bought $20 at a time, with no subscription.

terminalbash
# the local agent
pip install -U mlx-optiq
optiq code

# connect OptiQ Cloud, then type /boost in OptiQ Code before a hard task
optiq cloud login

OptiQ Code runs entirely on your Mac and needs no account. Boost is for the tasks a local model cannot finish on its own.