Free tools Windows power users keep installed
One-click scans. No signup required.
Maybe—but the evidence is a promising, narrow benchmark, not proof that any 2B model can reliably handle real work. In a 2026 report, Mingbird’s author says the harness scored 0.821 with gemma4:e2b (2B) on an 18-task artifact-based benchmark, versus 0.271 for goose, 0.017 for opencode, and 0.246 for agent-mini. The experiment used one Windows machine and one run per model–harness pairing. It is useful evidence to inspect and try to reproduce, not a general guarantee.
What Mingbird is—and what the benchmark asks
Mingbird is an open-source, local-first agent harness designed for Windows and Ollama, with both a graphical interface and a command-line interface. Its project documentation describes Apache-2.0 licensing, local inference by default, and support for Ollama and other OpenAI-compatible local servers. The repository lists Windows 10 and 11 as its main platform; Linux and macOS support is marked experimental. Mingbird’s GitHub repository provides its current documentation, code, and benchmark materials, while the releases page lists downloadable releases.
As an Amazon Associate I earn from qualifying purchases.
The benchmark is aimed at a practical question: can a harness help a small model complete tasks by managing tools and checking whether the requested deliverables actually exist? The project describes tasks involving code and tests, file organization, web research, data analysis, and MCP tools. The author’s framing includes familiar failure modes such as repeated tool calls, declaring a task complete while files are missing, or going silent. Those examples explain the motivation; they are not measurements of how often models fail in everyday use.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The scoring focuses on produced artifacts rather than judging transcripts alone. That is a useful distinction: a convincing explanation in a chat is not the same as a saved file, passing test, or completed task. But the result still depends on this benchmark’s chosen tasks, scoring rules, tools, and execution setup.
#1 Best Overall
What the 288-cell results say
In the author’s 2026 LRAB report, four harnesses were tested against four open models across 18 tasks, yielding 288 model–harness–task cells. The table gives the reported scores; higher is better within this benchmark. These are the project author’s measurements, not an independent evaluation.
| Model | Mingbird | goose | opencode | agent-mini |
|---|---|---|---|---|
| gemma4:e2b (2B) | 0.821 | 0.271 | 0.017 | 0.246 |
| qwen3.5:4b | 0.876 | 0.801 | 0.465 | 0.706 |
| gemma4:12b | 0.906 | 0.772 | 0.539 | 0.576 |
| ornith-1.5:35b | 0.941 | 0.679 | 0.896 | 0.092 |
| Overall | 0.886 | 0.631 | 0.479 | 0.405 |
The 2B row is the headline: Mingbird’s reported score is substantially higher than the other three harnesses’ scores for gemma4:e2b in this test. The lead is not uniform across all models or harnesses. For instance, opencode’s reported score with ornith-1.5:35b is 0.896, close to Mingbird’s 0.941, while the other pairings vary considerably. The overall row summarizes this particular collection of cells; it should not be read as a universal ranking.
The report also describes a separate small-model matrix using qwen3.5:2b: Mingbird scored 0.779, compared with 0.239 for goose, 0.096 for opencode, and 0.017 for agent-mini. The author additionally reports a τ²-bench comparison with Mingbird scores of 0.763 in retail, 0.740 in airline, and 1.000 in telecom. Those runs were performed by the project author; τ²-bench is identified as a Sierra Research benchmark, but this is not an independent evaluation of Mingbird.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHow the harness is intended to help
The project’s explanation emphasizes execution control and verification rather than a change to the underlying model. In the author’s account, three mechanisms are central:
- Finish gate: checks the original task against the deliverables before treating the work as complete.
- Verification feedback: passes actual task-test errors back into the process, giving the model information about what failed.
- Anti-loop handling: responds to repeated tool-call signatures or extended turns without output.
The article also describes a fixed 797-token factory prefill. The author says the per-mechanism ablation differences were smaller than single-cell variance, so the reported evidence does not establish reliable effect sizes for any one mechanism. The repository documents additional features, including automatic backups and rollback, bounded crawling of large files, chunked output feedback, and safeguards around destructive operations. These are project descriptions, not independently audited findings.
One task-level example illustrates why process efficiency and task success should be kept distinct. For task WF-08, the author reports 156 tool calls and zero files before a fix, versus 15 calls and three files after it, with the same bytes sent to the model. That is a single example, not an average across tasks or proof that the fix caused the difference in other settings.
What the experiment can—and cannot—establish
The LRAB campaign used one Windows machine, one trial per cell, a shared time budget, temperature 0, and thinking disabled. The author describes the results as directional rather than statistical. A single run cannot show how much scores would vary across repeated attempts, and a compact set of 18 tasks cannot represent every user’s workload.
- Task coverage: outcomes reflect the benchmark’s task set and artifact criteria, not every kind of coding, research, or file-management request.
- Hardware and backend: the campaign ran on one integrated-GPU computer with 32 GB of RAM, using the pinned backend and collection windows described in the project materials. Another machine, backend, or configuration may produce different outcomes.
- Harness comparison: the result compares particular implementations and settings, not every possible version or configuration of the named tools.
- Repeatability: one trial per cell provides no estimate of run-to-run variability or statistical confidence.
- Version: the headline LRAB matrix is tied to Mingbird v1.5.0, benchmark code at tag v1.5.0 and commit 1ee92d1. The project article described v1.8.2 as current when it was published; the repository’s README observed October 7, 2026 lists v2.1.2. The headline matrix was not run on v2.1.2.
The author makes the LRAB tasks, scoring code, raw per-cell data, and reproduction instructions public, which makes the result inspectable and gives readers a way to test it. Public materials improve reproducibility; they do not convert a one-machine, one-trial benchmark into independent confirmation. For an individual decision, the most relevant follow-up is to run representative tasks with the same model and setup the reader intends to use.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Trying it on a local computer
The repository lists integrated graphics and 16 GB of RAM as an entry tier for 2–4B models; the reported benchmark machine had 32 GB of RAM and integrated graphics. These figures describe project guidance and the author’s test setup, not a guarantee of speed or compatibility. Existing hardware may be enough to try local inference; the project article advises starting with a laptop already owned rather than assuming a purchase is necessary.
- Check platform and release: consult the repository README for current setup instructions and platform status, then use the releases page for available installers. Windows 10/11 is the documented main platform; Linux and macOS are experimental.
- Choose a local model server: the project documents Ollama and other OpenAI-compatible local servers. Follow the current README for the exact configuration supported by the release you install.
- Start with a bounded task: choose work with an observable result, such as creating a specified file or running a small test, and confirm the actual output rather than relying only on the agent’s completion message.
- Reproduce the published comparison only if needed: use the repository’s benchmark instructions, pinned code and backend details, and raw data. Record your own hardware, model, harness version, and settings so your result is not mistaken for the author’s benchmark.
Privacy and operational safety
Mingbird’s README says inference is local by default and that the application does not use an account, telemetry, update check, or analytics. The project also says web search and user-configured MCP servers can send traffic outside the computer; its offline toggle disables web tools and the optional cloud path. These are statements from the project documentation, not independently audited guarantees. Readers should review the current README’s description of online-mode endpoints and configure network access to match their needs.
Local inference does not make an agent harmless: the README warns that it reads and writes files and runs commands. Select a working directory deliberately, avoid exposing sensitive files unnecessarily, and review proposed destructive operations. A local model can still alter or delete data available to it.
How to judge whether the result matters for you
The benchmark is most useful as evidence that harness design may materially affect a small model’s performance on a defined set of tool-using tasks. It does not show that every 2B model can complete arbitrary work, nor that Mingbird will outperform alternatives on a different workload. Before relying on it, consider the dimensions that can change the outcome:
- Whether your tasks resemble the benchmark’s tasks and artifact checks.
- Which model family and size you plan to run.
- How the harness handles tools, verification, and recovery from errors.
- Your inference backend and computer.
- Whether you need offline operation, and which optional network-connected tools you enable.
- Whether your own repeated runs produce useful, verifiable deliverables.
The Mingbird README captures the project’s own qualification: “A 2B model will not rewrite your entire codebase in one shot — but it handles the bulk of everyday agent work, and when it cannot, it fails loudly instead of failing silently.” Treat that as the project’s characterization, not an independent finding. The benchmark supports a more limited conclusion: on the author’s 18-task setup, Mingbird performed strongly with the tested 2B model, and the public artifacts make the claim worth reproducing on tasks that matter to you.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
