Qwen3.8-27B-Pulsar

Qwen3.8-27B on a Mac at 156 tok/s writing code and 347 tok/s editing it (up to 477 tok/s; Pulsar 1.1.7), served by Pulsar.

Pulsar, a fork of Inco AI's open-source Splash engine, is the fastest engine we have measured for Qwen3.8-27B on Apple silicon (M5 Max). This repository holds everything Pulsar needs to serve the model: Inco AI's Splash package for Qwen3.8-27B (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

Qwen3.8-27B on an M5 Max (128 GB). Speeds are tokens per second (tok/s).

Test	Guesses ahead with	Other engine	Pulsar	Pulsar is
13 prompts, from short chat to a 32K-token agent task, greedy, Pulsar 1.1.5 (2026-10-09)			108.0
32K-token agent prompt, greedy, Pulsar 1.1.6 (2026-10-10; 1.1.5: 77.9)			83.3
vs lithos-metal 0.1.2 (official release), 10 prompts, greedy, Pulsar 1.1.7 (2026-10-14)	DSpark draft head	81.6	115.2	1.42×
vs Splash 1.3.0, 12 prompts, sampled, Pulsar 1.1.5 (2026-10-10)	DFlash 2 draft model	86.5	98.6	1.13×
vs MTPLX, 6 prompts, sampled, Pulsar 1.1.5 (2026-10-10)	MTP	64.9	125.0	1.93×
vs AX Engine, 5 prompts, greedy, Pulsar 1.1.7 (2026-10-14)	MTP	38.3	123.2	3.22×
vs MLX-LM, 6 prompts, sampled, Pulsar 1.1.7 (2026-10-14)	no draft model (its default)	31.4	129.5	4.12×
vs llama.cpp, 6 prompts, sampled, Pulsar 1.1.7 (2026-10-14)	no draft model (its default)	27.5	122.3	4.45×

Pulsar 1.1.7 is 1.42× faster than lithos-metal 0.1.2 (official release), and was faster in every session we ran and on every prompt. 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. Pulsar 1.1.6 writes 6.9% faster than 1.1.5 on the 32K-token agent prompt and reads that prompt in 40.1 s (1.1.5: 40.6 s). MTP is the model's own built-in guesser. Each comparison ran on the same Mac with the same prompts. Every session, prompt, setting and per-prompt number: BENCHMARKS.md.

Quick start

Apple silicon M3 or newer (measured on M5 Max and M5 Pro), macOS 26.4 or later, 24 GB of memory or more (24 GB Macs need one extra setting, see the README).

# 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/Qwen3.8-27B-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/                 Qwen3.8-27B, 4-bit, one packed file per layer   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
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/Qwen3.8-27B-Splash at revision 9d27070b71f7142c6b6025f03ac011d70a73cb48; manifest.json adds one record for the guesser list. Check every file with shasum -a 256 -c SHA256SUMS.
• head-ranked.u32 is Pulsar's guesser token list for Qwen3.8-27B (token ids only, no text): 2.4% more tok/s. The full model still checks every token.

Credits

• Qwen3.8-27B by the Qwen team, Apache-2.0.
• 4-bit conversion by mlx-community: mlx-community/Qwen3.8-27B-4bit, Apache-2.0.
• Splash and the DFlash 2 draft model by Inco AI: incoai/Qwen3.8-27B-Splash, incoai/Qwen3.8-27B-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.