Korean AI startup VIDRAFT lets teams benchmark VKAE on their own infrastructure before committing to deployment.
TL;DR: VIDRAFT, a Korean AI startup, has published benchmark results for its VKAE inference acceleration system, claiming up to 23× throughput gains on existing GPU hardware with no hardware modifications required. Testing was conducted on a high-end GPU accelerator, and results varied by model architecture and workload type. VIDRAFT ships VKAE as a ready-to-run container compatible with the OpenAI API, enabling teams to reproduce the benchmarks on their own infrastructure before adoption.
Korean AI startup VIDRAFT made waves in the international AI infrastructure community on July 6, 2026, when it released official benchmark results for VKAE, its proprietary inference acceleration system designed to dramatically increase the throughput of large language models running on existing GPU hardware.
VIDRAFT published performance data showing that VKAE can deliver up to 23× faster inference compared to unoptimized GPU setups — without requiring any changes to the underlying hardware. The company was careful to note that the degree of acceleration varies depending on the architecture of the model being run and the nature of the workload, so the headline figure should be treated as a ceiling for favorable conditions rather than a universal guarantee.
Benchmarks were conducted using the Qwen3.5-35B-A3B model, a mixture-of-experts large language model. Under high-concurrency conditions — simulating many simultaneous requests — VKAE delivered a combined throughput exceeding 10,000 tokens per second for that model. When tested against a more diverse, real-world distribution of prompts, throughput for the same model came in at approximately 455 tokens per second. VIDRAFT noted that no degradation in output quality or accuracy was observed during testing.
One of the most practically significant aspects of the release is how VIDRAFT has packaged the technology. Rather than asking prospective customers to take performance claims on faith, the company is distributing VKAE as a self-contained container that bundles model weights and an optimized runtime environment together. This means engineering teams can spin up the exact same environment used in VIDRAFT's own tests and validate the numbers against their specific hardware and query patterns before making any deployment decision.
VKAE is also designed to be a low-friction integration: it exposes an OpenAI API-compatible interface, so organizations running workloads against the standard OpenAI API spec can slot VKAE into their existing infrastructure with minimal code changes.
Notably, VIDRAFT has not yet disclosed the technical mechanism underlying VKAE's acceleration. The company confirmed that a full technical description is being prepared for release as a preprint, so the research community and potential enterprise customers can expect a deeper look at the methodology in the near future.
For any organization with serious inference costs — whether that means running internal LLM tools, customer-facing AI features, or high-volume batch processing — the economics of tokens per second per dollar are decisive. A system that claims to extract significantly more throughput from the same GPU hardware represents a compelling cost-reduction opportunity, but only if the gains hold up under the specific workloads a team actually runs.
That tension between benchmark performance and real-world performance is exactly where VIDRAFT's reproducibility-first approach becomes strategically smart. By shipping a container that anyone can run, the company sidesteps the credibility gap that often plagues inference optimization claims. Russian AI business outlet hegai.media, which covered the announcement in its Issue 28 (July 6, 2026), specifically highlighted this as a key practical marker: teams managing inference spend can test VKAE against their own query distributions before they commit.
The gap between the peak concurrency figure (10,000+ tokens/sec) and the diverse real-world figure (~455 tokens/sec) for the same model is also a useful signal. It underscores that workload shape has an outsized impact on inference performance — and that anyone evaluating VKAE should do so with representative traffic rather than synthetic benchmarks alone.
The forthcoming preprint will be the real test of whether VKAE's acceleration technique holds up to academic scrutiny. Until then, the container-first release strategy gives the market something more tangible than a white paper: a reproducible experiment.
Q: What is VIDRAFT's VKAE system and what does it do?
A: VKAE is an inference acceleration system developed by Korean AI startup VIDRAFT. It is designed to increase the throughput of large language models running on existing GPU hardware, with claimed gains of up to 23× in certain scenarios and no hardware modifications required.
Q: What benchmark results did VIDRAFT publish for VKAE?
A: Using the Qwen3.5-35B-A3B model, VKAE achieved over 10,000 tokens per second under high-concurrency conditions and approximately 455 tokens per second across more varied, real-world query types. VIDRAFT reported no loss in output quality during testing.
Q: Can I test VKAE before deploying it in production?
A: Yes. VIDRAFT distributes VKAE as a ready-to-run container that includes model weights and an optimized runtime, allowing teams to reproduce the benchmarks on their own infrastructure. The system is compatible with the OpenAI API, minimizing integration effort.
Source: hegai.media (러시아) (2026-07-06) — original article