DoubleChecks

What was claimed

NVIDIA just achieved 423 tokens/s/user on a 2.8T parameter model (1.68x faster) via new vLLM optimization with a 12,000 LoC megakernel

Our verdict

Accurate

SemiAnalysis reported 423 tokens per second per user for the 2.8-trillion-parameter Kimi K3 model.

Combined assessment15 sourcesChecked Oct 19, 2026

Why this verdict

  • The optimization involved a 12,000 LoC megakernel

    Misleading
    One assessment reached this conclusion; one differed·Claude: Verified

  • The performance result applies generally to a 2.8-trillion-parameter model.

    Misleading
    One assessment reached this conclusion

  • The achievement was 1.68x faster than previous performance

    Verified
    Two assessments reached this conclusion; one differed·ChatGPT: Can’t verify

  • NVIDIA achieved 423 tokens per second per user on a 2.8-trillion-parameter model.

    Verified
    Two assessments reached this conclusion; one differed·ChatGPT: Can’t verify

Evidence reviewed

15 sources

Perplexity, Claude and ChatGPT sources are combined here.

digg.comx.comx.comx.comx.comdigg.comx.comx.comithome.compost.smzdm.com

Individual assessments

Assessment summary

  • The core performance figures and vLLM optimization are supported by contemporaneous reports, but the wording overstates and simplifies the evidence.
  • The 12,000-line figure refers to roughly 12,800 added lines across many files, not necessarily one literal megakernel, and the 1.68x result is a specific benchmark rather than a universal speed claim.digg.com x.com

Claims reviewed

The optimization used a 12,000-line megakernel.

Misleading
The implementation used persistent, fused kernels and was described as a megakernel, but the roughly 12,800 lines were added across about 50 files rather than being one literal 12,000-line kernel.x.com

This claim may mislead readers without added context.

The performance result applies generally to a 2.8-trillion-parameter model.

Misleading
The figure is a benchmark for Kimi K3 under particular GB300 deployment and workload conditions. Results vary with GPU configuration, request count, context length, and speculative decoding.x.com

This claim may mislead readers without added context.

The result was 1.68x faster.

Verified
The reported comparison gave a 1.68x improvement on the stated benchmark setup and hardware.digg.com x.com

Assessment summary

  • All major claims in the statement are factually accurate and well-supported by recent sources.
  • The specific metrics (423 tokens/s/user, 2.8T parameter model, 1.68x speedup, 12,000 LoC megakernel) and the vLLM optimization context are all confirmed by multiple credible sources from October 2026.

Claims reviewed

NVIDIA achieved 423 tokens/s/user on a 2.8T parameter model

Verified
Multiple sources confirm NVIDIA reached 423 tokens per second per user on the 2.8-trillion-parameter Kimi K3 model digg.com source unavailable source unavailable.

The achievement was 1.68x faster than previous performance

Verified
Sources explicitly state the new optimization made the 2.8T parameter model 1.68 times faster on the same hardware digg.com source unavailable.

This was achieved via new vLLM optimization

Verified
The optimization is described as an upstream vLLM optimization implemented through aggressive Megakernel fusion digg.com source unavailable.

Assessment summary

  • The statement mixes verifiable vLLM/Kimi K3 performance reports with details that are either reported differently across sources or not officially attributed to NVIDIA alone.
  • Some sources report similar throughput gains and mention a megakernel implementation, but the exact 423 tok/s, 1.68× figure and the "12,000 LoC" characterization are reported in secondary coverage and differ from official vLLM posts.
  • Because parts are supported and parts are inconsistent or unverified, the overall claim is ambiguous.

Claims reviewed

The gain was achieved via a new vLLM optimization (a megakernel) implemented as ~12,000 lines of code (LoC).

Misleading
vLLM documentation and blogs describe megakernel/mono-kernel and other fused-kernel optimizations that increase throughput, and independent articles mention a "~12k LoC megakernel." However, line-count as a causal or official metric is not confirmed in primary vLLM/NVIDIA technical posts, making the precise "12,000 LoC" claim a secondary-source characterization rather than an official technical claim. vllm.ai source unavailable.

This claim may mislead readers without added context.

NVIDIA just achieved 423 tokens/s/user on a 2.8T parameter model.

Can’t verify
vLLM-related benchmarks for a 2.8T model (Kimi K3) report high single-user throughputs (e.g., 118 tok/s baseline, 370 tok/s with DSpark in one vLLM post) and community reports claim 423 tok/s, but no single authoritative NVIDIA or vLLM post unequivocally states "NVIDIA just achieved 423 tok/s/user." vllm.ai vllm.ai.

This performance is 1.68× faster.

Can’t verify
Some coverage computes a ~1.68× uplift when comparing 252 tok/s baseline to 423 tok/s, but the vLLM official blog reports other speedups (for example 3.14× in a different configuration to 370 tok/s). The 1.68× ratio appears in secondary reports but is not confirmed by an authoritative primary source. vllm.ai vllm.ai.

Check your own claim

Paste any statement, headline, or AI answer. Independent assessments check it and compare sources.

Share this result