Paris99 commited on
Commit
8285c0d
·
verified ·
1 Parent(s): 57d7fef

Card: title mentions the bundled DSpark speculative decoding

Browse files
Files changed (1) hide show
  1. README.md +3 -2
README.md CHANGED
@@ -14,7 +14,7 @@ tags:
14
  - speculative-decoding
15
  ---
16
 
17
- # Inkling 2 Mini for Apple Silicon (MLX, 4/8-bit, lossless)
18
 
19
  [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini) (revision `6e7c824`) for MLX on
20
  Apple Silicon, with its DSpark speculative-decoding draft bundled (`mtp.safetensors`) and its vision encoder included.
@@ -28,6 +28,7 @@ Inkling 2 Mini is Thinking Machines' looped mixture-of-experts model: 26.4B para
28
  | Build | Size | Same next token as Thinking Machines' code | KL vs their code | GSM8K (50 questions) | Peak memory, 8K-token prompt |
29
  |---|---|---|---|---|---|
30
  | [netprojects/Inkling-2-Mini-MLX-4-8bit](https://huggingface.co/netprojects/Inkling-2-Mini-MLX-4-8bit) (recommended) | 20.2 GB + 0.64 GB draft | 98.6% | 0.0018 | 47/50 | 27.8 GB |
 
31
 
32
  "Thinking Machines' code" is their Transformers implementation of Inkling 2 Mini, run on the released weights; the
33
  columns are explained under "What was checked".
@@ -100,7 +101,7 @@ harnesses and settings; a selection of their table):
100
 
101
  ## What was checked
102
 
103
- On a MacBook Pro M5 Max (128 GB), 20 November 2026, with MLX 0.32.4 and [inkling2-mlx](https://github.com/internet-projects/inkling2-mlx) v0.1.0.
104
 
105
  - **Against Thinking Machines' code.** Their Transformers implementation (pull request #51050) ran on the CPU with the
106
  released weights, on 12 short chat texts (488 answer positions). This build gave the same next token at
 
14
  - speculative-decoding
15
  ---
16
 
17
+ # Inkling 2 Mini for Apple Silicon (MLX, 4/8-bit, lossless, with DSpark speculative decoding)
18
 
19
  [thinkingmachines/Inkling-2-Mini](https://huggingface.co/thinkingmachines/Inkling-2-Mini) (revision `6e7c824`) for MLX on
20
  Apple Silicon, with its DSpark speculative-decoding draft bundled (`mtp.safetensors`) and its vision encoder included.
 
28
  | Build | Size | Same next token as Thinking Machines' code | KL vs their code | GSM8K (50 questions) | Peak memory, 8K-token prompt |
29
  |---|---|---|---|---|---|
30
  | [netprojects/Inkling-2-Mini-MLX-4-8bit](https://huggingface.co/netprojects/Inkling-2-Mini-MLX-4-8bit) (recommended) | 20.2 GB + 0.64 GB draft | 98.6% | 0.0018 | 47/50 | 27.8 GB |
31
+ | [netprojects/Inkling-2-Mini-MLX-3bit](https://huggingface.co/netprojects/Inkling-2-Mini-MLX-3bit) | 15.5 GB + 0.64 GB draft | 94.9% | 0.0430 | 49/50 | 21.8 GB |
32
 
33
  "Thinking Machines' code" is their Transformers implementation of Inkling 2 Mini, run on the released weights; the
34
  columns are explained under "What was checked".
 
101
 
102
  ## What was checked
103
 
104
+ On a MacBook Pro M5 Max (128 GB), 21 November 2026, with MLX 0.32.4 and [inkling2-mlx](https://github.com/internet-projects/inkling2-mlx) v0.1.0.
105
 
106
  - **Against Thinking Machines' code.** Their Transformers implementation (pull request #51050) ran on the CPU with the
107
  released weights, on 12 short chat texts (488 answer positions). This build gave the same next token at