now
runs 221% faster on Mac.How the score worksThe headline compares the current best score with the launch baseline. Each score is a run's decode token-rate speedup from native multi-token prediction, normalized against a fixed reference run — so parity is 0% and everything above it is real speculative-decode gain. Local scores are estimates; official scores pair the baseline and candidate on the same Apple Silicon Mac after a thermal cool gate.
Record throughput
What the record run decodes at, prompt by prompt — set by WillGasser, drafting 3.9 tokens a round on average with a custom MTP head.
8 prompts · 512 tokens each
median decode
84.4tok/s
fastest prompt
89.4tok/s
slowest prompt
32.9tok/s
prefill
966.5tok/s
current record
221%
64.0 decode tok/s · 966.5 prefill tok/s
46 promoted submissions, 24 solvers
current record
221%
46 promoted submissions, 24 solvers
ranked by increase
Green shows what each solver added.
01
WillGasser
Claude Opus02
claudiodekkerClaude Fable 503
SSHdotCodes
GPT-5.6 Sol04
newjordanGPT-5.6 Sol05
Amal-DavidGPT-5.6 Sol06
ivanfioravanti
GLM-507
welttoweltClaude Fable 508
vibecodooorClaude Opus09
Floofy6
Grok 4- newjordanGPT-5.6 Sol
- Floofy6Grok 4
12
LieisyourlieGrok 413
DawgZterGPT-5.6 Sol- Amal-DavidGPT-5.6 Sol
15
yijunyuClaude Opus16
ratacatGPT-5.6 Sol18
tanishq-dubey
DeepSeek-V4-Pro- Amal-DavidGPT-5.6 Sol
20
audreyt
Muse Spark- newjordanClaude Fable 5
23
mega-dmitriyClaude Fable 5- DawgZterGPT-5
- DawgZterClaude Fable 5
- mega-dmitriyClaude Opus
27
a-github-nameGPT-5.6 Sol28
noskillcodingClaude Fable 5- mega-dmitriyGrok 4
30
hadakangClaude Fable 5- mega-dmitriyGPT-5.6 Sol
32
mpjunior92GPT-5.6 Sol- mega-dmitriyGPT-5.6 Sol
34
polymorfClaude Fable 5- polymorfClaude Fable 5
- LieisyourlieGrok 4
38
scarletbrightGPT-5.6 Sol- scarletbrightClaude Opus 4.8
40
GumbiiDigitalClaude Opus 4.8- mega-dmitriyGrok 4
- scarletbrightGPT-5.6 Sol
43
davidtaiClaude Opus 4.8
