Pulsar 发布：在 M5 Max 上以 148 tok/s 运行 Qwen3.8-27B 写代码

The fastest inference engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon.
Qwen3.8-27B: up to 452 tok/s editing code · 1.36× lithos-metal (greedy, 10 prompts) · 1.11× faster than Splash 1.3.0
measured on an M5 Max · every token still checked by the full model

_{lithos-metal's own launch-video prompt, same Mac: Pulsar 1.1.3 writes 3,000 tokens in 20.4 s (147.8 tok/s), lithos-metal 0.1.2 in 27.0 s (111.8 tok/s). Video}

_{Editing a pasted 210-line file: the first 1,884 tokens take Pulsar 1.1.3 8.1 s and lithos-metal 0.1.2 16.5 s (321 vs 142 tok/s). Video}

_{Qwen3.8-27B on an M5 Max (128 GB, macOS 27.2), greedy, one server at a time (2026-10-09). lithos-metal runs NVIDIA's NVFP4 checkpoint with its DSpark draft head. Real-time replays of measured token timings.}

_{Same prompt, same 708-token answer: Pulsar 3.97 s · Splash 1.3.0 4.07 s · MLX-LM 23.1 s. Video}

_{Building a particle galaxy app in one shot: Pulsar writes its 3,124 tokens in 20.9 s, Splash 1.3.0 takes 22.1 s for the same number of tokens, then the galaxy Pulsar wrote runs. Video}

_{Editing a pasted 210-line file: Pulsar 8.2 s · Splash 1.3.0 11.7 s · MLX-LM 64.5 s. Video}

_{Qwen3.8-27B 4-bit on an M5 Max, one engine's server per pane, greedy answers (2026-10-07). Real-time replays of measured token timings.}

Jump to: How fast? · Requirements · Quick start · What's different · Switches · All the numbers

Pulsar (formerly fastkernel) runs the Qwen3.8-27B and Qwen3.6-35B-A3B AI models on your own Mac. Chat with them in your browser, or connect a coding agent or any other app.

Pulsar 1.1.4 is built on Splash 1.3.0, the latest Splash. On an M5 Max it writes answers 1.11× faster than Splash 1.3.0, and a repeated prompt starts in 0.1 s from its prompt cache. It brings everything new in Splash 1.3.0. A small draft model guesses ahead, and the full model checks every token before it keeps it.

M5 Max (128 GB)

Speed you see (Qwen3.8-27B 4-bit, Pulsar 1.1.0, greedy, measured from the stream, 2026-10-07):

Peak is the most tokens in any one second; average runs from the first token to the last. Each row is one run, the first (nothing cached); the first token includes reading the prompt, here a 1,825-token pasted file.

Pulsar 1.1.3 vs lithos-metal 0.1.2, Qwen3.8-27B (2026-10-09):

On our 128 GB M5 Max running macOS 27.2, Pulsar 1.1.3 delivered 1.36× the client-observed decode throughput of lithos-metal 0.1.2 across these 10 prompts (geometric mean; one ABBA session, greedy decoding, thinking off, output capped at 1,024 tokens). The engines used different quantized checkpoints and drafters; task-quality parity was not evaluated.

Decode tok/s, mean of two runs per prompt. Same Mac, one server at a time in the order Pulsar, lithos-metal, lithos-metal, Pulsar, each started with its README command plus port and context size, after at least 120 s of warm-up. lithos-metal runs NVIDIA's NVFP4 checkpoint with its DSpark draft head; Pulsar runs the 4-bit Splash package with DFlash2. Settings and limits: docs/BENCHMARKS.md.

Pulsar 1.1.0 vs other engines (2026-10-07):

Speeds are in tok/s (tokens per second). A token is a word or part of a word. Each pair ran in one session, taking turns on the same prompts: 12 prompts against Splash 1.3.0 (chat, math, code, a code file, a 32K-token agent task, a multilingual prompt and two long-context prompts), the 6 standard prompts against the others (5 greedy prompts against AX Engine). In chat, Pulsar 1.1.0 writes 76.7 tok/s, 1.12× Splash 1.3.0's 68.5.

Sent again, a 2,247-token prompt starts in 0.1 s: Pulsar 1.1.0 reuses its prompt cache.

Also: Qwen3.6-35B-A3B (MoE)

MoE means "mixture of experts". The model has many small expert blocks and uses only a few of them for each token. That makes it fast for its size. On Qwen3.6-35B-A3B, Pulsar 1.1.0 writes 268.1 tok/s sampled and 299.6 greedy, against Splash 1.3.0's 260.9 and 276.1 on the same 6 prompts (2026-10-07).

Full numbers and how we measured: docs/BENCHMARKS.md

Requirements

Quick start

Both ways below start a server at http://127.0.0.1:8000. It speaks the OpenAI and Anthropic APIs, so apps built for either one can use it. To chat, open that address in your browser. The first run downloads the model (17.4 GB).

24 GB Mac? First, run sudo sysctl iogpu.wired_limit_mb=20480. It lets the GPU use 20 GB until you restart the Mac. Then put SPLASH_TEXT_ONLY=1 in front of the serve command. It skips the part of the model that reads images.

For Qwen3.6-35B-A3B, use incoai/Qwen3.6-35B-A3B-Splash instead of incoai/Qwen3.8-27B-Splash.

Prebuilt download (no Xcode)

Download pulsar-1.1.4-macos-arm64.tar.gz from Releases. Then unpack it and start the server:

tar -xzf pulsar-1.1.4-macos-arm64.tar.gz && cd pulsar
SPLASH_DRAFT_HEAD_IDS=$PWD/data/head-ranked.u32 ./pulsar serve --model incoai/Qwen3.8-27B-Splash

Got the file through a browser or AirDrop? Then macOS blocks it until you run /usr/bin/xattr -dr com.apple.quarantine . once in the pulsar folder. Run it before the serve command.

The first run also sets up its Python packages, which takes 20 seconds.

Build from source

git clone https://github.com/loopai-hq/pulsar && cd pulsar
make install MODEL=incoai/Qwen3.8-27B-Splash
SPLASH_DRAFT_HEAD_IDS=$PWD/data/head-ranked.u32 ./pulsar serve --model incoai/Qwen3.8-27B-Splash

Use it with a coding agent

A coding agent is an AI tool that reads and edits your code, like OpenCode, Claude Code or Codex. Leave the server running. Open a second terminal in the pulsar folder and run:

./pulsar opencode    # or: ./pulsar claude / ./pulsar codex / ./pulsar hermes

The agent opens already connected to Pulsar. Other apps can use the OpenAI API at http://127.0.0.1:8000/v1 or the Anthropic API at http://127.0.0.1:8000, with the model incoai/Qwen3.8-27B-Splash.

Skill for coding agents

A skill is an instruction file that a coding agent reads. .claude/skills/pulsar teaches an agent to install, start, connect and stop Pulsar. Claude Code and OpenCode find it when you open them in this folder. To use it from any folder:

cp -R .claude/skills/pulsar ~/.claude/skills/    # Claude Code and OpenCode
cp -R .claude/skills/pulsar ~/.agents/skills/    # Codex

What's different from Splash

• Built on Splash 1.3.0. Pulsar 1.1.4 brings its speedups onto the latest Splash, with its prompt reading and prompt cache: a repeated prompt starts in 0.1 s.
• Same model, same checking. A small draft model guesses the next few tokens. The full model checks every guess and keeps only the ones it agrees with.
• Guesses from your prompt. Some answers repeat your input, like an edit to a file you pasted. There the engine takes its guesses straight from your prompt. The full model can then keep many tokens in one step.
• Faster GPU code. New GPU code, hand-tuned for Apple silicon, does the model's math in less time.
• Text-only mode. SPLASH_TEXT_ONLY=1 skips the image weights, which leaves more memory for the context on smaller Macs.
• New engine code. 6,191 new lines of C++ and Metal, Apple's GPU language, on top of Splash 1.3.0. They include 38 new GPU kernels, small programs that run on the GPU. What each part does and what it gained: docs/WHATS-INSIDE.md. Every change has a switch: docs/SWITCHES.md.

Credits & license

Pulsar is licensed under the Apache License 2.0 (LICENSE, NOTICE). If you build on it, keep the NOTICE file and credit Pulsar.

Thank you to Inco AI for Splash, the Apache-2.0 engine Pulsar is built on. Pulsar 1.1.4 is built on Splash 1.3.0. Splash's own README: docs/SPLASH-README.md. Third-party code that Splash ships keeps its own license: THIRD_PARTY_NOTICES. Files we changed from Splash say "Modified by meowkernels."; the full list, including data files that can't hold a comment, is in docs/CHANGED-FILES.md.