GitHub - loopai-hq/pulsar: Pulsar (formerly fastkernel): the fastest inference engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon

Qwen3.8-27B on your Mac: 153 tok/s writing code, 344 tok/s editing it.

The fastest engine we have measured for Qwen3.8-27B on Apple silicon (M5 Max), a fork of Inco AI's Splash 1.3.0.
Up to 477 tok/s editing code · 1.44× faster than lithos-metal 0.1.2 (official release) · 1.13× Splash 1.3.0
every token still checked by the full model

_{lithos-metal's own launch-video prompt, same Mac: Pulsar 1.1.5 writes 3,000 tokens in 19.7 s (153 tok/s), lithos-metal 0.1.2 in 25.3 s (120 tok/s). Video}

_{Editing a pasted 210-line file: Pulsar 1.1.5 finishes its 1,894-token answer in 7.3 s (344 tok/s, up to 477), lithos-metal 0.1.2 its 1,884 tokens in 16.1 s (151 tok/s). Pulsar takes most of this answer straight from the pasted file (prompt lookup). Video}

_{Qwen3.8-27B on an M5 Max (128 GB, macOS 27.2), greedy, thinking off, one request at a time (2026-10-10). lithos-metal 0.1.2 is the official release, running NVIDIA's checkpoint (NVFP4 MLP, FP8 attention) 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 (MLX-LM without a draft model, its default). 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 (MLX-LM without a draft model, its default). 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? · What's new in 1.1.7 · Requirements · Quick start · What's different · Models · Switches · All the numbers

Pulsar (formerly fastkernel) is a fork of Inco AI's Splash 1.3.0. It uses Splash's model packages and DFlash 2 draft models unchanged and adds new engine code: GPU kernels, drafting and prompt lookup. It 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.

A model writes its answer one token at a time; a token is a word or part of a word. Pulsar pairs the full model with a small guesser, a draft model that proposes the next few tokens. The full model checks every guess in one step and keeps only the ones it agrees with, so you get the full model's answers, faster.

On an M5 Max, Pulsar 1.1.5 averages 108.0 tok/s across 13 prompts, from short chat to a 32K-token agent task (greedy, 2026-10-09), and a repeated prompt starts in 0.1 s from its prompt cache.

M5 Max (128 GB)

Speed you see (Qwen3.8-27B 4-bit, Pulsar 1.1.5, greedy, measured from the stream, 2026-10-10):

Peak is the most tokens in any one second; average runs from the first token to the last. The edit row is the first run (nothing cached), so its first token includes reading the 1,825-token pasted file; the app row is the median of three runs.

Pulsar 1.1.5 vs lithos-metal 0.1.2, Qwen3.8-27B (2026-10-10):

On our 128 GB M5 Max running macOS 27.2, Pulsar 1.1.5 delivered 1.44× the client-observed decode throughput of lithos-metal 0.1.2, its official release (geometric mean over 10 prompts; one ABBA session, greedy decoding, thinking off, output capped at 1,024 tokens). Decode tok/s, mean of two runs per prompt:

The first token arrives in 0.22 s on Pulsar and 0.39 s on lithos-metal (median). lithos-metal runs NVIDIA's checkpoint, which keeps attention in FP8: it has a 17.6 GB logical target-weight payload per verification pass, against 14.4 GB for Pulsar's package, and part of Pulsar's lead comes from that smaller payload. The drafters differ (DSpark vs DFlash 2), and answer quality was not compared.

All sessions: docs/BENCHMARKS.md.

Pulsar vs other engines:

Speeds are in tok/s (tokens per second). All rows are Pulsar 1.1.5 (2026-10-10). MTP is the model's own built-in guesser. 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 1.1.5 writes 74.9 tok/s, 1.13× Splash 1.3.0's 66.4.

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 1.1.5 writes 264.3 tok/s on average across 13 prompts, from short chat to a 32K-token agent task (greedy, 2026-10-10).

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

What's new in 1.1.7

Qwen3.6-35B-A3B writes faster, and several requests at once run faster. Answers match 1.1.6. Measured against the previous release on the same M5 Max, taking turns on the same prompts (greedy, 2026-10-10):

Both 35B runs used the guesser token list on both sides. The 13 new prompts were never used to tune Pulsar. The new kernels apply on the 40-core M5 Max; other Macs run 1.1.6's.

1.1.6: long prompts write faster, 6.9% on a 32K-token agent task with Qwen3.8-27B (77.9 → 83.3 tok/s).

Download and source. Pulsar 1.1.7 ships as a ready-to-run download, the latest Pulsar. This repository holds the Apache-2.0 source of Pulsar 1.1.4, which builds and runs as before.

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, run ./pulsar serve --model incoai/Qwen3.6-35B-A3B-Splash, without SPLASH_DRAFT_HEAD_IDS.

Download Pulsar 1.1.7 (latest, no Xcode)

Download pulsar-1.1.7-macos-arm64.tar.gz from Releases. Then unpack it and start the server:

tar -xzf pulsar-1.1.7-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 Pulsar 1.1.4 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 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. The guesser proposes 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.
• A lighter guesser. The guesser picks from a shorter list of the tokens the model writes most (data/head-ranked.u32), so each guess costs less.
• Faster GPU code. New GPU kernels, tuned by measurement on an M5 Max, do 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. The 1.1.4 source adds 6,191 lines of C++ and Metal, Apple's GPU language, to Splash 1.3.0, including 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.

Models on Hugging Face

Each model has a page with its speed table and quick start:

• loopai-hq/Qwen3.8-27B-Pulsar
• loopai-hq/Qwen3.6-35B-A3B-Pulsar

They carry Inco AI's Splash model packages unchanged, with full credit, plus Pulsar's guesser token list for Qwen3.8-27B.

Credits & license

Pulsar is licensed under the Apache License 2.0 (LICENSE, NOTICE). The download includes LICENSE, NOTICE and THIRD_PARTY_NOTICES. If you build on Pulsar, keep the NOTICE file and credit Pulsar. Pulsar was called fastkernel until version 1.1.3.

Thank you to Inco AI for Splash, the Apache-2.0 engine Pulsar is built on, and for the DFlash 2 draft models and Splash model packages Pulsar runs. Pulsar 1.1.7 is built on Splash 1.3.0. Splash's own README: docs/SPLASH-README.md. Thank you also to the Qwen team for the Qwen models and to mlx-community for the 4-bit conversions. Third-party code that Splash ships keeps its own license: THIRD_PARTY_NOTICES. Files we changed from Splash say "Modified by Pulsar."; the full list, including data files that can't hold a comment, is in docs/CHANGED-FILES.md.