For coding with one of the OpenCode free models, start with Muse Spark 1.3 Free or Space Bunny Free. Both passed every hidden test on the feature and bug-fixing tasks and produced the cleanest refactors. For writing and planning documents, Big Pickle led our AI-judged (blind) scores. These are first-round results from our own test of all nine free models.

We ran the test on 15 and 16 October 2026 with OpenCode 2.0.29 and our own free-tier account, at a reported cost of $0. Each model received the same ten tasks: five coding and five writing. Coding was scored by hidden automated tests the models never saw. Writing and one architecture task were scored blind by an AI judge. The data is still interim: 177 of 270 planned executions are recorded and 124 are scored. Writing and architecture have one run per model. Treat everything below as round one, not a final ranking.

Quick picks

Scroll horizontally to read all columns.

If you needUseWhyWatch out for
An everyday coding agentMuse Spark 1.3 FreePassed every hidden test on the coupon service, bug-fix and feature tasks in all three runs; 4.1-minute median coding run; 30 of 30 runs finished without a platform errorIts id contains “contributor”. Whether prompts are used by the provider is unverified, so check the terms before sending confidential code
Coding without that question markSpace Bunny FreeHighest coding total (97.2), well-structured refactors, 84 on the architecture document (AI-judged, blind)About 7 minutes per coding task
Fast iterationsCat JamFastest median coding run at 2.8 minutesWeaker refactoring (90) and architecture (59, AI-judged, blind); first model to hit the free-tier cap
Articles, rewrites and planning documentsBig PickleBest writing total (76.6), best Romanian article (84) and best architecture document (85), all AI-judged (blind)Slower than Muse Spark and Cat Jam on coding
Nothing agenticAvoid Nemotron 3.5 Lightning FreeLowest coding total (57.9) and lowest writing total (40.1, AI-judged, blind)—

These picks are our editorial judgement from the first round of data, not a guarantee of how each model will behave on your project.

Coding results: hidden automated tests

Every model got the same five coding tasks: build a coupon and discount service with an API in TypeScript, fix a buggy ticketing library, refactor messy legacy code without changing its output, design the architecture of a clinic-booking SaaS, and implement a feature from a one-page issue in an existing repository.

Scroll horizontally to read all columns.

ModelCoding total /100Coupon serviceBug fixingRefactorArchitecture (AI-judged, blind)Feature taskMedian coding run (min)
Space Bunny Free97.210010097841007.0
Muse Spark 1.3 Free95.9100100100731004.1
Big Pickle93.99910076851005.4
Cat Jam91.61001009059952.8
LongCat 2.5 Preview Free87.599100556410014.8
Step 5 Preview Free78.69910060010023.7
Nemotron 3 Ultra Free77.495977409125.0
Nemotron 3.5 Lightning Free57.9686671284111.7
MiMo-V2.6-Flash Freenot yet complete99100——9514.8

Scores are means over the runs completed so far, which is one to three per task. The coding total weights the coupon service at 30%, bug fixing at 25% and the other three tasks at 15% each. Because the architecture task is AI-judged, 15% of the coding total is an opinion rather than a test result. MiMo has no coding total yet because its refactor and architecture runs had not finished.

Bug fixing no longer separates most models

The ticketing library contained 12 planted bugs. Users had reported only seven of them, so the other five could only be found by reading the code. Seven of the nine models fixed all 12 in every run so far, including the five unreported ones. The exceptions were Nemotron 3 Ultra, which missed the waitlist bug in one of two runs, and Nemotron 3.5 Lightning. If you choose between free models, a simple bug-fix task tells you little about which is better.

Refactoring separates them most

The legacy file started with 22 any types, 66 var declarations, three module-level mutable variables and a 94-line function. All 240 output comparisons had to keep passing.

Muse Spark scored 99–100 in all three runs. It kept every output check, removed all the any types and var declarations, and split the code so that its longest function was 40–45 lines. Space Bunny scored 94–100 and spread the code across 12–14 small files. Cat Jam preserved behaviour but left functions of 94–103 lines, so it scored 90. LongCat, Step 5 and one Big Pickle run hit the time limit without changing the code, which scored 55–60.

For refactoring that should leave code easier to maintain, Muse Spark and Space Bunny were the only models that did it consistently. Passing tests is still not the same as being ready for users; run a pre-launch checklist for an AI-built app before you ship what any model writes.

The feature task

The autonomous task asked each model to implement a feature from a one-page issue. It was then checked against 17 hidden tests. Seven models passed all 17 in every scored run. Nemotron 3 Ultra passed 16 of 17. Nemotron 3.5 Lightning passed 8, 3 and 4 in its three runs.

Where Nemotron 3.5 Lightning fell short

Nemotron 3.5 Lightning Free was clearly the weakest coding model in this round. On the coupon service it passed 33% of the concurrency tests and 20% of the API tests. In none of its three bug-fixing runs did it fix the “lowest price” sorting bug or the async race condition. Its architecture document scored 28 (AI-judged, blind). It did finish all 30 of its runs, so the problem is the quality of the work, not reliability.

Writing results: AI-judged (blind)

The five writing tasks were a technical article about WooCommerce hosting for 10,000 products, an evidence-only comparison of three fictional hosts, a 700-word short story with a twist, a rewrite of a weak 900-word article without adding facts, and a persuasive product article in Romanian. An AI judge scored anonymous texts in random order. Each model has one run per task.

Scroll horizontally to read all columns.

ModelWriting total (AI-judged, blind)Technical articleEvidence comparisonShort storyRewriteRomanian article
Big Pickle76.67876707684
MiMo-V2.6-Flash Free75.78173638082
Muse Spark 1.3 Free74.87677607982
Space Bunny Free69.17971537963
LongCat 2.5 Preview Free68.37474577759
Step 5 Preview Free68.07755757856
Cat Jam65.16773466871
Nemotron 3 Ultra Free54.15462445854
Nemotron 3.5 Lightning Free40.14643275826

Each task score combines accuracy (25%), depth (25%), naturalness (20%), structure and usefulness (20%) and instruction following (10%).

The top three are close. Big Pickle had the best total and the best Romanian article. MiMo had the best technical article and rewrite. Muse Spark had the best evidence comparison and was within two points of the leader overall. Step 5 is the interesting split: it wrote the best short story (75) but scored only 55 on the evidence comparison, where accuracy against the evidence mattered. Only Nemotron 3.5 Lightning scored lower on that task.

With a single run per model and an AI judge, these scores are an informed opinion. A difference of one or two points between the top three should not decide anything on its own.

Most models ignored length limits

Length briefs were often missed. Some technical articles ran past 2,000 words against a 1,200–1,800 brief. One story reached 1,351 words against a 630–770 brief. Cat Jam went the other way, writing 356 words for a story of about 700, which is why it scored 46. If length matters for your page, check it yourself rather than trusting the model to follow the brief, and verify sources the same way you would check citations in an AI research answer.

Speed and reliability

Median minutes per coding run varied by about nine times:

Scroll horizontally to read all columns.

ModelMedian coding run (min)Runs with an attributable outcome
Cat Jam2.826 of 29
Muse Spark 1.3 Free4.130 of 30
Big Pickle5.413 of 15
Space Bunny Free7.021 of 22
Nemotron 3.5 Lightning Free11.730 of 30
LongCat 2.5 Preview Free14.811 of 13
MiMo-V2.6-Flash Free14.89 of 12
Step 5 Preview Free23.711 of 12
Nemotron 3 Ultra Free25.012 of 14

The slow group (LongCat, MiMo, Step 5 and Nemotron 3 Ultra) usually reached correct code eventually. These were also the models with timeouts and provider stalls. Nemotron 3 Ultra timed out in 7 of 14 runs and Step 5 in 4 of 12. Both timed out on the architecture task without producing a document, which is why they score 0 there. Nemotron 3 Ultra and Step 5 also repeatedly returned 504 Upstream idle timeout mid-run, with some streams stalling for 10–20 minutes. We classified those stalls as provider failures and retried them; they did not count against the models’ scores.

In practice, if you cannot wait 15–25 minutes per task or tolerate a stalled stream, stay with Cat Jam, Muse Spark, Big Pickle or Space Bunny.

Free-tier limits and when to switch models

The free tier is capped per model, not per account. When a model reaches its cap, OpenCode returns FreeUsageLimitError with HTTP 429 and the message “Rate limit exceeded. Please try again later.” The error does not say when the limit resets.

Under our heavy load, with nine models running agentic tasks in parallel, the first model to hit the cap was Cat Jam after 1 hour 19 minutes. Five more models hit it between 2 hours 22 minutes and 4 hours 19 minutes after the start. Muse Spark, Nemotron 3 Ultra and Nemotron 3.5 Lightning never hit it during the test. Blocked models stayed blocked for at least 2.7 to 6.3 hours. The test machine then slept overnight, so we did not observe the exact reset, but the models answered normally the next morning. OpenCode does not publish the exact quotas, and we did not measure them.

Two more rules showed up in practice. The free tier only accepts requests made from inside OpenCode: a request with every tool disabled was rejected with HTTP 403 and the message “OpenCode’s free tier can only be used from within OpenCode.” And because the caps are per model, switching to another free model usually worked when one was blocked.

A sensible fallback order from these results:

  • Coding: Muse Spark 1.3 Free, then Space Bunny Free, then Big Pickle or Cat Jam. Muse Spark never hit the cap in our test, which makes it a good model to switch to.
  • Writing: Big Pickle, then MiMo-V2.6-Flash Free or Muse Spark 1.3 Free. Their AI-judged (blind) writing totals were within two points of each other.

How to run a specific free model in OpenCode

OpenCode is an open-source AI coding agent available as a terminal CLI and a desktop app; the version 2 documentation is at opencode.ai/v2/docs. Run opencode models to list what your account can use. Ours listed 40 models, most of them paid. The free ones have cost 0 in the account’s model metadata, and most have -free in their id.

To run one model on one task without the interactive interface:

opencode run -m opencode/space-bunny-free "your prompt"

Adding --format json returns a machine-readable event stream that includes per-step token counts and cost.

Scroll horizontally to read all columns.

ModelOpenCode idContext window (tokens)
Big Pickleopencode/big-pickle200,000
Cat Jamopencode/cat-jam-free262,144
LongCat 2.5 Preview Freeopencode/longcat-2.5-preview-free1,000,000
MiMo-V2.6-Flash Freeopencode/mimo-v2.6-flash-free200,000
Muse Spark 1.3 Freeopencode/muse-spark-1.3-contributor-free1,048,576
Nemotron 3 Ultra Freeopencode/nemotron-3-ultra-free1,000,000
Nemotron 3.5 Lightning Freeopencode/nemotron-3.5-lightning-free262,144
Space Bunny Freeopencode/space-bunny-free1,048,576
Step 5 Preview Freeopencode/step-5-preview-free1,000,000

Note the Muse Spark id. A separately listed muse-spark-1.3 is paid. The word “contributor” in the free id suggests that prompts may be used by the provider, but we did not verify this. Read the model’s terms before you send it code or data you would not want shared.

How we tested

  • Tasks: ten per model, identical for all nine: the five coding and five writing tasks described above.
  • Coding scores came from hidden automated tests the models never saw: 51 tests for the coupon service, 21 for the bug fixes, 240 output comparisons for the refactor and 17 for the feature task. All were written and validated before any model ran.
  • Writing and architecture scores came from a blind AI judge (Claude), which saw only anonymous texts in random order. These are labelled AI-judged throughout.
  • Fairness: every run used the same prompt, the same starting files, the same tools, the same time limit (15–30 minutes depending on the task) and the same step budget, in a fresh session with no web access.
  • Safety: commands from the models ran in an isolated container with no internet access and no access to personal files.
  • Settings: each model ran with its default variant; none used a higher reasoning setting.
  • Cost: OpenCode reported $0 for all runs.
  • Status: interim. 177 of 270 planned executions are recorded and 124 are scored. Third repetitions were still running because of free-tier limits. The test design, hidden tests and analysis were prepared with Claude Code.

Disclosure

We ran this test ourselves on 15–16 October 2026 with OpenCode 2.0.29 and our own free-tier account. Neither OpenCode nor any model provider sponsored or endorsed it. Coding results come from hidden automated tests the models never saw. Writing quality was scored blind by an AI judge, so treat those scores as an informed opinion. Free models change often, so re-test on your own tasks before relying on these results.

Sources and checking

Product terms can change. These are the sources checked for this article; follow the links to verify current details before you buy.