How to use from
Docker Model Runner
docker model run hf.co/husky8/Inkling-2-Fast-Computer-Use
Quick Links

Inkling 2 Fast Browser Use

A fresh six-case run of this two-row checkpoint on the MacBook Pro measured 2.39脳 end to end, compared with 2.56脳 in the first run. On the repeat the edit passed 6/6, the original passed 4/6, and median time on the four common successes fell from 16.62 s to 6.96 s. The same filmed case measured 1.03脳 on one RTX 5090. All attempts and fresh video

A separate four-row checkpoint passed WebArena search task 274 on three different RTX 5090s, with observed wall-time speedups of 1.52脳, 1.54脳 and 1.55脳. Search results 路 Four-row checkpoint

This project measures browser tasks in Brave. The Mac form-and-cart checkpoint and the RTX 5090 WebArena checkpoint are separate edits. Ten WebArena results.

This checkpoint swaps two rows in Inkling 2 Mini鈥檚 output layer. On six paired Brave runs, the model filled a form and added the requested quantity to a cart in a median of 7.93 seconds, down from 20.34 seconds. Both checkpoints completed all six runs with identical commands. The test ran on a MacBook Pro using MLX and text descriptions of visible controls. Screenshot-based control and other browsers are untested.

BrowserGym

The original passed 11/20 cases and the edit passed 6/20. Both passed 5. The original alone passed 6, the edit alone passed 1, and both failed 8. Across the 5 joint successes, median completion time was 19.86 s original and 3.84 s edited. Four paired seeds per environment. Task scores and receipts

Download

hf download husky8/Inkling-2-Fast-Computer-Use --local-dir inkling-browser

In an Inkling-compatible computer-use agent, select husky8/Inkling-2-Fast-Computer-Use as the model. The agent still supplies observations and executes actions. The full checkpoint is about 20 GB. Load the directory with a runtime that supports the original Inkling 2 Mini format. Benchmark, video and runnable example

Weight change

Rows 200004 and 200008 of model.llm.unembed.weight exchanged places. These rows score the answer and thinking channel markers. On this task the edit eliminated generated thinking tokens. Both checkpoints used greedy decoding and reasoning effort none. In a separate development run, prefilling the original checkpoint鈥檚 answer channel also produced a similar time. General reasoning and visual accuracy after this edit have not been evaluated.

Base revision ed40e7da625f594fc9c4990f5f66be24873c22d7. Edited shard SHA-256 d713d16818032b5ac0d06b07e72d4dc7135d57737e3224216d09b51afe59ab6e.

Independent modification of Thinking Machines Lab鈥檚 Inkling 2 Mini, distributed under Apache 2.0. Original model card 路 Model acceptable use policy 路 Modification notice

Downloads last month
-
Safetensors
Model size
26B params
Tensor type
F32
路
BF16
路
F8_E4M3
路
U8
路
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for husky8/Inkling-2-Fast-Computer-Use

Quantized
(32)
this model