Qwen3.6-35B-A3B-Pulsar

Qwen3.6-35B-A3B on a Mac at 264.3 tok/s, served by Pulsar.

Pulsar is a fork of Inco AI's open-source Splash engine. Qwen3.6-35B-A3B is a mixture-of-experts (MoE) model: 35B parameters in many small expert blocks, of which only about 3B work on each token, so it is fast for its size. This repository holds Inco AI's Splash package for it (the 4-bit model, its DFlash 2 draft model, the vision encoder and the tokenizer), unchanged, plus Pulsar's guesser token list. A model writes one token at a time; a token is a word or part of a word. A small guesser, the draft model, proposes the next few tokens and the full model checks every one before it keeps it, so you get the full model's answers, faster.

Speed

On an M5 Max (128 GB), greedy, measured on Inco AI's Splash package without the guesser list (2026-10-10):

• Pulsar 1.1.5 writes Qwen3.6-35B-A3B at 264.3 tokens per second (tok/s) on average across 13 prompts, from short chat to a 32K-token agent task.
• On the 32K-token agent prompt, Pulsar 1.1.6 writes 139.9 tok/s, 3.6% faster than 1.1.5's 135.1.

With the guesser token list, on the same Mac:

• Pulsar 1.1.7 writes 282.7 tok/s across the 13 prompts, 4.8% faster than 1.1.6's 269.7 (both with the guesser token list, greedy, 2026-10-10).

Since 2026-10-10 this package also ships Pulsar's guesser token list (draft/head-ranked.u32); answers are unchanged (greedy, 12 prompts identical with and without it).

Installed before 2026-10-10? Pulsar keeps an installed model as it is. To pick up the list, remove the model link (the weights stay in the cache) and serve again; it downloads about 1 MB:

rm "$HOME/Library/Application Support/Splash/models/loopai-hq/Qwen3.6-35B-A3B-Pulsar"
./pulsar serve --model loopai-hq/Qwen3.6-35B-A3B-Pulsar

More numbers: BENCHMARKS.md.

Quick start

Apple silicon M3 or newer (measured on M5 Max), macOS 26.4 or later.

# Download pulsar-1.1.7-macos-arm64.tar.gz from https://github.com/loopai-hq/pulsar/releases/latest, then:
tar -xzf pulsar-1.1.7-macos-arm64.tar.gz && cd pulsar
./pulsar serve --model loopai-hq/Qwen3.6-35B-A3B-Pulsar

Open http://127.0.0.1:8000 to chat. The server speaks the OpenAI and Anthropic APIs, and ./pulsar opencode, ./pulsar claude or ./pulsar codex opens your coding agent already connected. More in the README.

What's inside

target/                 Qwen3.6-35B-A3B, 4-bit (8-bit expert routers)   Inco AI's Splash package, unchanged
draft/                  DFlash 2 draft model                            Inco AI's Splash package, unchanged
draft/head-ranked.u32   Pulsar's guesser token list                     Pulsar
vision/                 bf16 vision encoder                             Inco AI's Splash package, unchanged
tokenizer/              tokenizer and chat template                     Inco AI's Splash package, unchanged
manifest.json           size and SHA-256 of every package file          Inco AI's, plus one record for the list
layout.json             map of every packed file                        Inco AI's Splash package, unchanged
LICENSE                 Apache-2.0                                      Inco AI's, unchanged
NOTICE                  credits and what Pulsar changed                 Pulsar
provenance.json         source repos, revisions and licenses            Pulsar
SHA256SUMS              SHA-256 of every file here                      Pulsar

The Splash package files are byte-for-byte copies of incoai/Qwen3.6-35B-A3B-Splash at revision 0f4714b2db37b5f3c42a10de07281e74f88e4adc; manifest.json adds one record for the guesser list. Check every file with shasum -a 256 -c SHA256SUMS.

Credits

• Qwen3.6-35B-A3B by the Qwen team, Apache-2.0.
• 4-bit conversion by mlx-community: mlx-community/Qwen3.6-35B-A3B-4bit, Apache-2.0.
• Splash and the DFlash 2 draft model by Inco AI: incoai/Qwen3.6-35B-A3B-Splash, incoai/Qwen3.6-35B-A3B-DFlash2 and the Splash engine, Apache-2.0. Thank you to the Splash team at Inco AI for the engine and the model packages Pulsar stands on.
• Pulsar and its guesser token list, Apache-2.0: github.com/loopai-hq/pulsar.

License

Apache-2.0, like every component above.