Serving got 2.8x faster with zero new hardware. vllm's september 13 optimization of kimi k3 cut…
serving got 2.8x faster with zero new hardware. vllm's september 13 optimization of kimi k3 cut latency 56-60% and first-token time up to 85% just by tuning the engine.
the model didn't change, the serving code did.
Context
vLLM's blog of 13 September 2026 compares v0.27.1 against a main-branch commit on Kimi K3, with an 8K input and 1K output workload on a B300 node, TP8 and eight-token DSpark speculation, at concurrency 1, 4 and 16. Average latency fell from 12.37 to 5.30 seconds (minus 57.2 percent), 23.67 to 10.50 (minus 55.6) and 55.90 to 22.17 (minus 60.3); average first-token time fell from 2262.9 to 376.3 milliseconds (minus 83.4 percent), 2314.9 to 640.5 (minus 72.3) and 7601.1 to 1121.0 (minus 85.3); throughput rose 120, 150 and 180.6 percent.
It is vendor-reported, on one workload and one hardware tier, and not independent. The 2.8x is throughput and not latency, and the latency gain is about 2.3 to 2.5 times by my arithmetic from the table. The comparison is against a main-branch commit and not a tagged release. Zero new hardware holds as the same B300 node on both sides, but just by tuning the engine simplifies many software changes such as scheduler, prefix cache and kernels. Other figures in the blog use different configurations and are not merged here.
Related work
- Earlier note on the same vLLM blog ↗Same blog and source.
- Earlier note on vLLM and Kimi K3 ↗Same optimization, framed as 2.8 times faster.
Watch next
- The release tag containing the commit and independent reruns.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 21 September 2026 at 16:22 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →