Splash

Apple M3 or newer, macOS 26.4 or later, Homebrew, and 36 GB of unified memory (48 GB or more recommended).

brew install incoai/tap/splash
splash serve --model incoai/Qwen3.8-27B-Splash

The first run downloads and verifies the model package, checks available memory, and starts serving on 127.0.0.1:8000.

Once it prints Ready, leave this terminal open. Open in your browser, or run an installed coding agent from another terminal:

splash opencode    # or: splash claude / splash codex / splash hermes

Press Ctrl+C in the server terminal to stop Splash.

Use the API

Splash speaks OpenAI Chat Completions (/v1/chat/completions), OpenAI Responses (/v1/responses), and Anthropic Messages (/v1/messages), all with streaming, tool calls, JSON Schema output, images, and inline PDFs. /tokenize and /apply-template return token IDs and the rendered prompt without running the model.

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "incoai/Qwen3.8-27B-Splash",
    "messages": [{"role": "user", "content": "Explain speculative decoding in one sentence."}]
  }'

model is optional. If you set it, it must match the package you served. Reasoning is on by default. "reasoning_effort": "none" turns it off, and Qwen3.8-27B also takes low, medium, and xhigh.

/v1/judgments and /v1/systemone provide scoring without generation. See judgment contracts for details.

Models

| Package (--model) | Contents | Download | | --- | --- | ---: | | incoai/Qwen3.8-27B-Splash | Qwen3.8-27B, 4-bit, with its DFlash 2 draft | 17.4 GB | | incoai/Qwen3.6-35B-A3B-Splash | Qwen3.6-35B-A3B, 4-bit, with its DFlash 2 draft | 20.9 GB |

--model takes any owner/repo that holds a Splash package, a format DEVELOPMENT.md describes. Plain MLX or Transformers checkpoints do not work. Private repositories need HF_TOKEN. Packages download into the Hugging Face cache, and brew upgrade splash keeps them, along with model links and agent sessions.

For a custom model download location, see model cache.

Settings

There is no config file. The server binds 127.0.0.1:8000 by default. Context supports up to the model’s native 256K window; usable capacity depends on available memory.

splash serve accepts these optional flags:

• --host: HTTP bind address. Default: 127.0.0.1.
• --port: HTTP port. Defaults to SPLASH_PORT or 8000.
• --max-memory: ceiling on Metal allocations, e.g. 28G. Default: auto.
• --max-context: context limit, up to 256K, e.g. 100K. Default: auto.
• --max-image-pixels: maximum resized pixels per image. Default: 4,194,304.
• --allowed-host: extra HTTP Host name to accept, not a bind address. Repeatable.
• --api-key: require this key on API requests, as a bearer token or

• --no-webui: turn off the chat page.

Authentication is off by default. Set SPLASH_API_KEY in the shell that runs splash serve and in the shell that runs an agent, and both sides use it. Health and readiness probes stay public.

For LAN access and multiple servers, see server configuration.

Experimental cache offloading: PR #3.

Performance

Measured on an M5 Pro (16-core GPU, 48 GB): selected SPEED-Bench coding prompts over HTTP, a 1,024-token output limit, reasoning on (medium for the 27B). The ratio in each cell is against the next-fastest engine we measured.

| Metric | Qwen3.6-35B-A3B | Qwen3.8-27B | | --- | ---: | ---: | | Decode · short prompt | 210 tok/s (1.7×) | 74 tok/s (2.0×) | | Prefill · 32K prompt | 2,011 tok/s (1.3×) | 363 tok/s (1.2×) | | Cached time to first token · 32K replay | 123 ms (6.6×) | 282 ms (7.3×) | | Aggregate decode · 4 concurrent short prompts | 357 tok/s (2.0×) | 170 tok/s (3.9×) |

Splash led on every measure at every prompt length we tested, and the lead grows with load: 3.8× at four concurrent 32K requests on the 35B. The launch post has the method and the full comparison against oMLX, Lily, uzu, and Ollama.

Design

The runtime, scheduler, cache, and API are shared. Everything else is rebuilt per model:

• A draft trained for the model. Speculative decoding is the decode path in

• Kernels compiled for exact shapes. Fused Metal kernels, written and tuned

• A memory plan computed for this machine. Context, KV capacity, and batch

The launch post covers the design in depth.

More

• DEVELOPMENT.md: building from source, tests, model packages,

• Apache-2.0, see LICENSE. Model weights keep their own licenses.