FlashMap's speedup is the authors'
FlashMap's paper reports up to 6.14 times speedup on GSM8K with Qwen3-8B and about 7 percent over matched DFlash. Those figures are the authors' table.
The short version
The authors'. The arXiv paper and GitHub repo report up to 6.14 times on GSM8K with Qwen3-8B and about 7 percent average gain over a matched DFlash baseline. MarkTechPost summarized the same table.
The paper and the repo
FlashMap appears on arXiv as 2610.24365 with a public GitHub repository under al-jshen/flashmap. The authors report speculative-decoding draft speedups, including up to 6.14 times on GSM8K with Qwen3-8B and about 7 percent average improvement over a matched DFlash setup. MarkTechPost summarized the same claims on October 27. Author tables are not third-party audits. OCR-2's benchmark brief used the same habit: the repo's card is the card. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.
Who is in scope
Researchers and engineers who read speculative-decoding papers and who may try the public code. Product teams that never run draft models are outside the dependency. Naming the model and the benchmark is enough. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.OCR-2.
Genie 4 is the other AI filing
Genie 4 is a DeepMind preview with coming-soon product access. FlashMap is an efficiency paper. Both sit on the artificial-intelligence desk today. They do not share a release channel. Asked directly, Does this page reprint training recipes? No. The filing is the authors' reported numbers and the public paper and repo links. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Genie 4.
What this page will not paste
Training hyperparameters, loss schedules, and copy-paste recipes stay in the paper. This brief names the speedup claims and points to arXiv and GitHub. Shepherds as reviewers remains the human-oversight note beside lab scorecards. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Shepherds are still the reviewers.
D3 Ultra stays a scorecard
D3 Ultra's card was a different leaderboard claim. FlashMap's card is speculative decoding. Linking them only keeps AI measurement filings distinct. Asked directly, What peak GSM8K speedup do the authors report? Up to 6.14 times with Qwen3-8B on GSM8K, in the paper's table. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.D3 Ultra.
Terra stays an addendum
Terra's warning-rate addendum was a different lab document. FlashMap does not inherit that label. The short answer this page will stand on is this: The authors'. The arXiv paper and GitHub repo report up to 6.14 times on GSM8K with Qwen3-8B and about 7 percent average gain over a matched DFlash baseline. MarkTechPost summarized the same table. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Terra warning rate.
MarkTechPost as second account
MarkTechPost is an outlet summary of the authors' work. When the outlet paraphrases, the paper remains the primary for the 6.14 times and 7 percent lines. Asked directly, Is this Genie 4? No. Genie 4 is DeepMind's world-model preview. FlashMap is a speculative-decoding efficiency paper. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.
Simplismart was a round
Simplismart's 14 million dollar round was a financing brief. FlashMap is a paper. Distributed cloud and AI desks stay separate. Asked directly, What is FlashMap? A research method and open repository for faster draft models in speculative decoding, posted on arXiv as 2610.24365. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Simplismart.
The record
The record is arXiv 2610.24365, the FlashMap GitHub repository, and MarkTechPost. Speedups are the authors'. No training recipe on this page. Asked directly, How do they compare to DFlash? They report about 7 percent average improvement over a matched DFlash baseline in their experiments. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
What is FlashMap?
A research method and open repository for faster draft models in speculative decoding, posted on arXiv as 2610.24365.
What peak GSM8K speedup do the authors report?
Up to 6.14 times with Qwen3-8B on GSM8K, in the paper's table.
How do they compare to DFlash?
They report about 7 percent average improvement over a matched DFlash baseline in their experiments.
Is this Genie 4?
No. Genie 4 is DeepMind's world-model preview. FlashMap is a speculative-decoding efficiency paper.
Does this page reprint training recipes?
No. The filing is the authors' reported numbers and the public paper and repo links.