mrkeyoor.comAsk

incoai/splashtested 05 Oct 2026commit 4c11cbf

splash review

Splash merits a supported-Mac trial when its documented models and local serving interfaces fit your workflow.

  • Install 26 s
  • Build 6 s
220/293passed, run exited non-zero
  • 49 packages
Self-Hostedevaluationupdated 31 Oct 2026
+273stars / 7d

Splash is a local inference engine for Apple silicon. Splash runs coding agents and OpenAI or Anthropic compatible applications on one Mac. Consider it for a supported-Mac trial, but distinguish its documented capabilities from our October 5, 2026 development run at commit 4c11cbf. That run used Debian and did not validate Mac inference.

Verdict

Splash merits a supported-Mac trial when its documented models and local serving interfaces fit your workflow. Our October 5, 2026 run at commit 4c11cbf recorded successful installation and build-command outcomes in Debian, followed by a nonzero test outcome. It did not validate native inference or either supported installation route. Check compatibility and try your actual agent task before adopting it.

We ran it

Lab card: what happened when we ran splash
Lab card from our sandbox run
Screenshot of splash (inco.ai/blog/splash)
inco.ai/blog/splash
Install✓ · 26s49 packages · 81 MB
Build✓ · 6s
Tests✗ · 157s220 passed · 26 failed · 20 skipped · 47 errors of 293 (pytest)
Known vulns0(pip-audit)
Repo505 files~127,966 lines of source · 8.4 MB · 2 CI workflows · tests dir

Answers from our run

Does splash build from source?

Dependencies installed in 26 seconds (49 packages), and the build succeeded in 6 seconds. We cloned commit 4c11cbf into a clean Debian container with 3 CPUs and no project-specific setup.

Do splash's tests pass?

No. The project's own test command finished in 157 seconds and reported 220 passed, 26 failed, 20 skipped, 47 collection errors (pytest).

Does splash have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use splash?

Readers seeking a verified Linux or Windows deployment.

What are the alternatives to splash?

llama.cpp, MLX LM, Ollama. Splash merits a supported-Mac trial when its documented models and local serving interfaces fit your workflow.

Setup2/5Substantial documented constraints: Apple M3 or newer, macOS 26.4 or later, compatible weights and additional model downloads. Neither supported installation route was tested. The Debian development-test result is excluded from this judgment.
Docs5/5Under the visible documentation rubric, the supplied README gives unusually complete detail for a supported-Mac evaluation: installation, formats, APIs, memory settings and benchmark conditions. Source-build details remain unverified because the supplied development-guide URL returned 404.
Community3/5Last push on 2026-10-06 and 2,281 stars.Editor’s note: The supplied release notes show issue-linked maintenance but do not establish contributor breadth, participation or community support. They cannot support a numeric community judgment.
Maturity3/5A bounded offering under the visible rubric: release 1.2.1 documents concrete serving fixes and remaining sleep-related limits. Our run did not verify those fixes or production reliability.

How we score

Who it’s for

Developers with an Apple M3 or newer running macOS 26.4 or later who want to evaluate local inference.
Coding-agent users whose model family and weight format meet Splash's documented compatibility rules.
Contributors interpreting the recorded Debian development-command failures.

Who it’s NOT for

Readers seeking a verified Linux or Windows deployment.
Buyers needing independently measured inference speed, output quality or coding-agent reliability.
Users assuming every GGUF or MLX checkpoint is compatible.
Operators treating localhost binding or a zero advisory count as a security assessment.

Setup reality

Our October 5, 2026 run used commit 4c11cbf, with full SHA 4c11cbfea2ecb1bab41764b46a1195975e9c2017, in the dev subdirectory of a Debian Bookworm Python sandbox with 3 CPUs and 8 GB RAM. Installation returned zero in 26 seconds and recorded 49 packages with an 81 MB footprint. The build step returned zero in 6 seconds. The test command returned exit code 1 after 157 seconds.

The stored parser summaries report 220 passed, 26 failed, 20 skipped and 47 errors, with a stored total of 293. These figures do not establish exhaustive unique test coverage and have not been reconciled against complete logs. Skipped tests are not failures.

The supplied current README offers a desktop app and a Homebrew command-line route. Both require Apple M3 or newer and macOS 26.4 or later. The first serve downloads the model and its matching draft. Neither supported installation route nor native inference was tested in our sandbox.

Should you try Splash on your Mac?

Splash runs coding agents and OpenAI or Anthropic compatible applications on one Mac. Consider a trial when your hardware, model family and weight format meet its documented requirements. The supplied current README offers a desktop app and a Homebrew command-line route. Both require Apple M3 or newer and macOS 26.4 or later.

The current README documents streaming, tool calls, JSON Schema output, base64 images and PDFs, and a browser chat page. These are project-documented capabilities. Our sandbox did not verify those workloads or successful coding-agent task completion.

The first serve downloads the target model and its matching draft, so leave disk space for the weights. Agent launchers use installed clients. A launcher alone does not establish that the agent will complete your coding task.

What the Debian run established

Our run took place on October 5, 2026, at commit 4c11cbf, with full SHA 4c11cbfea2ecb1bab41764b46a1195975e9c2017. It used the dev subdirectory in a Debian Bookworm Python sandbox with 3 CPUs and 8 GB RAM.

Installation returned zero in 26 seconds and recorded 49 packages with an 81 MB footprint. The build step returned zero in 6 seconds. The test command returned exit code 1 after 157 seconds.

The stored parser summaries report 220 passed, 26 failed, 20 skipped and 47 errors, with a stored total of 293. These summaries do not establish exhaustive unique test coverage. No new execution or complete-log reconciliation was performed. The skipped tests should not be counted as failures.

The raw tail identifies several problems: image tests could not import PIL; makefile tests could not find /usr/bin/lockf; crash-replay tests reported that the replay engine reached EOF before ReadyEvent; and a build-configuration test returned an unexpected status. Another makefile subcase encountered a missing temporary environment path. The tail does not establish the cause of every failure or error.

The pytest footer reports 151.62 seconds and 2,830 passing subtests. Its timing scope differs from the harness's 157-second command duration. Those subtests are not additional unique tests to add to the stored total.

The successful build-command outcome does not establish that Metal kernels compiled, a model loaded or inference worked. The desktop app, Homebrew installation, Mac deployment, coding-agent completion, throughput and output quality remain untested by this run.

The recorded pip-audit advisory total is zero. Dependency scan coverage is unknown: the record supplies no scannedPackages value, and its packages field of zero does not establish coverage. This was not a security review.

The MrKeyoor testing methodology explains the publisher's command records and sandbox limitations. The supplied page describes a later procedure and does not establish which methodology version governed this October 5 run.

Match compatibility to the snapshot

The README at the tested commit listed Qwen3.8-27B, Qwen3.6-35B-A3B and Ternary Bonsai 2. Its main 4-bit examples required at least 36 GB of unified memory, with 48 GB recommended; smaller GGUF variants could run on 24 GB Macs. Those requirements describe that snapshot.

The supplied current README also lists Ornith-1.5-122B, Qwen3.8-Flash-Next and Flint. It documents affine MLX weights from 2 to 8 bits, including mixed precision, plus mxfp4 and nvfp4. The earlier narrow family list and blanket MLX 4-bit restriction should therefore not guide a current installation decision. Current documentation also does not establish what worked at our tested commit.

Check the family and format rules before downloading weights. The current README excludes whole-BF16 and F16 models and does not promise arbitrary checkpoint compatibility. Its performance tables are project-reported benchmarks on specified Macs and models; our sandbox provides no independent speed comparison.

Decide how the server will be accessed

The README says the server listens on localhost without authentication by default. That describes binding and authentication behavior, not a security guarantee. Consult the current server configuration before exposing it beyond the Mac.

The supplied development-guide URL returned 404. It cannot support the earlier detailed LAN requirements, endpoint exemptions or source-build requirements, so those assertions are omitted.

The current README documents limits for Metal memory, context and SSD cache, plus optional cache persistence. Defaults differ between the supplied snapshots, so use documentation matching the release you install.

Alternatives for different requirements

llama.cpp documents Apple silicon, x86 and several GPU backends. Evaluate it when hardware breadth is the deciding requirement. Its README also documents an OpenAI-compatible server. Our run does not establish which engine is faster for your workload.

MLX LM is a Python package for generating text and fine-tuning large language models on Apple silicon with MLX. Choose it for that generation or fine-tuning workflow; those capabilities alone do not establish equivalence with Splash's agent integrations.

Ollama documents installation routes for macOS, Windows and Linux. Evaluate it when those installation choices matter. It also documents a REST API for running and managing models. No alternative's documented integration establishes successful completion of your agent task.

What the ratings mean

These are editorial judgments about the supplied evidence, not benchmark results. For assessed dimensions, the five-point rubric is: 1 means little practical support is established; 2 means substantial documented constraints; 3 means a bounded offering with concrete capabilities and limits; 4 means broad documented coverage with material gaps; and 5 means unusually complete documentation for the decision covered here. An unassessed dimension has no numeric score. The ratings do not measure security or support response times.

Setup ease is 2 solely because the supported routes require Apple M3 or newer, macOS 26.4 or later, a compatible model and additional model downloads. Supported installation ease is untested. The Debian development-test failure does not determine this score.

Documentation is 5 for the supplied README's detailed treatment of installation, model formats, APIs, memory settings and benchmark conditions. This judgment concerns the supported-Mac evaluation covered here; the inaccessible development guide leaves source-build details unverified.

Community is unassessed. The supplied release notes document issue-linked fixes, but do not establish contributor breadth, participation or community support.

Maturity is 3 for a bounded offering with documented serving fixes and explicit remaining limits. The Splash 1.2.1 release notes describe fixes for tool-call arguments, requests spanning Mac sleep and configurable idle memory release. They exclude deep standby lasting hours and driver-reported GPU failures from the sleep fix. Release commit 1a13b75 is distinct from our tested commit 4c11cbf, and our run did not verify those fixes.

Before adopting Splash, try the release on your intended Mac with the exact model, format and coding-agent task you plan to use.

Alternatives

ProjectWhat it isPick it when
llama.cpp gh↗A C/C++ inference engine with documented CPU and GPU backends and an OpenAI-compatible server.Evaluate it when hardware breadth is the deciding requirement; verify the backend and model you intend to use.
MLX LMA Python package for text generation and fine-tuning on Apple silicon with MLX.Choose it for an MLX generation or fine-tuning workflow; verify serving and agent requirements separately.
Ollama gh↗A model runner with a REST API for running and managing models.Evaluate it when documented macOS, Windows and Linux installation routes fit your deployment needs.

What people are saying

  1. [velocity-scout] incoai/splash
  2. [hackernews] Splash-free urinals (2025)

Sources

  1. Splash README
  2. Splash 1.2.1 release
  3. Splash development guide
  4. MLX 8-bit support request

More self-hosted reviews