hao-ai-lab/jetspec is an LLM inference project that extends speculative decoding (a technique that first drafts candidate tokens and then confirms them through verification, advancing in a single pass) to tree drafts. The latest commit (673267e), merged on October 30, 2026, is PR #33, and it touched documentation only, not code. It reorganized the README section by section and newly added AGENTS.md to the methods and engine directories. The change amounts to 207 added lines and 219 deleted lines across 8 markdown files.
The README was rewritten with top links (the paper arXiv 2610.19792, the project webpage, Hugging Face models), an overview, supported models, a quick start, evaluation, and code. The JetSpec and JetSpec++ descriptions, which had been hidden behind collapsible blocks, were expanded, and the code section links to the methods implementation (a torch reference) and the serving engine (which runs on 2 GPUs) via their respective READMEs. Rather than copying the paper's Table 3 over again, the evaluation section points toward the bench/jetspec_plus_plus/reproduce_paper_results.md guide. This guide explains the host file, the protocol, and other baselines.
A sentence stating the conditions on the numbers was added to the quick start. The output throughput is a value obtained by processing a single short prompt on the new engine, so it comes with the hint that, to see a fixed baseline, you should use bench's speedup.py. The README states up to 7.45x versus autoregressive decoding and 1.54x versus DSpark on 2 GB200s, with up to 1,361 tokens/sec, and records that across 7 benchmarks the accepted length averages 8.94 tokens per round at temperature 0 (DSpark: 6.89).
That said, this commit is entirely documentation cleanup. The performance figures are values carried over from the paper's Table 3, not results re-measured by this change, and the decoding behavior is unchanged. The entry point and directory guidance are easier to read, but there is still no independent verification of the speed claims, and the GB200 and B200 hardware assumed by the README is required. It is closer to a cleanup worth following along with than something new to try.