ConwayResearch/husky-flash

Husky Flash: Woof 1.0 4B for Husky, with the Flash draft

Woof is the Pareto-frontier model that runs Underdog on your Mac and iPhone, even without wifi. This is Woof packed for Husky, Underdog's model-specific inference engine for Apple silicon, with Flash on: the trained draft that lets Husky check several tokens per read of the weights.

Husky is built around Woof's exact shapes instead of around any model. On an M5 Max it is faster than Apple's MLX on all sixteen tasks we measure: 1.8 to 3.9× on edits, 3.6 to 5.5× to the first word on a continued conversation, and with Flash on, 1.3 to 4.5× across the board. Same weights, same answers. The numbers, the method and the receipts are at husky.underdog.ai.

What is in the package

• target/: Woof's 4-bit weights laid out in the order Husky's kernels read them. A repack, not a requantization: no value changes.
• draft/: the Flash draft, round 4. One small layer that reads Woof's own hidden states from five of its layers and proposes the next seven tokens; trained by self-distillation on Woof's replies to 139,000 conversations: Underdog's own prompt shapes (edits, JSON, tables, HTML and markdown UI, tool calls, mail, code, summaries, browser steps, follow-ups) and public chat, instruction and code sets. On writing it lands about 2 tokens a step; on code and structured output, 4 to 5.
• manifest.json: the model's geometry, which the engine reads instead of assuming.
• tokenizer/: Woof's tokenizer.

How to run it

Download Underdog for the Mac and Husky runs Woof for you. In Underdog, the Woof card on the Models page has the engine switch: Apple MLX, Husky, or Husky with Flash on.

Developers can serve this package directly with the Husky server from the Underdog Greyhound repository:

husky serve --model ConwayResearch/husky-flash

Measured on this package (M5 Max, greedy, medians of three, quiet windows)

Prompt	MLX tok/s	Husky	Flash on
Short email	151	164	260
Reply to a thread	163	170	217
Tone rewrite	162	168	269
Meeting notes to to-dos	163	175	234
Call summary	162	166	211
Question over a document	159	193	210
Invoice to JSON	164	208	496
Data to a table	163	188	282
Write a function	155	170	545
CSV to a table	158	287	462
Fix typos in a paragraph	159	463	535
Rename a SQL column	158	487	547
Add a field to a JSON file	157	611	672
Function edit	163	614	730

First token on a continued conversation, both engines from their caches: 29 to 39 ms on Husky, 137 to 177 ms on MLX.

Licence

Woof's licence applies to these weights, as for ConwayResearch/woof-1.0-4B.