Browser task speedups

Best verified speedup for each of the 11 selected Mac tasks across released checkpoints. Both models passed each run shown. Edited times link to the exact weights. Filmed timings

Task Original Inkling 2 Mini Edited Inkling 2 Mini Speedup
Latest completed order total 路 WebArena 190 45.89 s 5.13 s 8.95脳
Latest canceled order total 路 WebArena 188 45.75 s 6.44 s 7.10脳
Complete an autocomplete field 路 MiniWoB++ seed 103 19.86 s 3.84 s 5.18脳
Reply to an email 路 MiniWoB++ seed 104 26.01 s 5.98 s 4.35脳
Search for Xbox 路 WebArena 275 20.43 s 4.96 s 4.12脳
Fill a form, set quantity and add a dress to the cart 25.52 s 6.35 s 4.02脳
Fill and submit a form 11.19 s 2.79 s 4.02脳
Search for green tea bags 路 WebArena 278 24.29 s 6.12 s 3.97脳
Search for iPhone 13 batteries 路 WebArena 277 19.38 s 5.88 s 3.30脳
Search for a USB Wi-Fi adapter 路 WebArena 274 14.46 s 4.65 s 3.11脳
Search for Switch accessories 路 WebArena 276 14.39 s 5.15 s 2.79脳

On the completed-order lookup, inference fell from 43.03 s to 2.84 s. The original model generated 1,509 thinking tokens. The edit generated zero. Both made the same clicks and returned $65.32.

Weight-space programming. We edited Inkling 2 Mini鈥檚 output matrix (model.llm.unembed.weight) without training. The order checkpoints swap the answer-text and thinking rows (200004 and 200008) and adjust the click-command and end-of-generation rows (34 and 200006). The completed-order edit changed 9,192 BF16 values. The autocomplete and email runs used the two-row swap available on main. Weight audit

Method. All 11 featured tasks ran on an M5 Max MacBook Pro with MLX. Both models read visible text and numbered controls in Brave. Each pair used the same task, initial page, prompt, greedy decoding, reasoning effort (none) and cursor timing. Task time includes inference and browser actions after warmup. We developed the order edits on order-lookup tasks. WebArena runs 路 MiniWoB++ runs 路 Search runs 路 Reproduction instructions

Both checkpoints peaked at 20.8 GiB of active MLX memory during a separate replay of the filmed order lookup on the 128 GiB Mac. Memory measurements

Base model by Thinking Machines Lab. Inkling 2 Mini 路 Apache 2.0 and 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-Mini-Fast-Browser-Use

Quantized
(96)
this model