confiron.Open app

Configuration is boring.
Let Confiron do it for you.

Discover the right model, optimize inference settings, and benchmark performance across serving engines—all backed by measurable results.

Explore the technology Enter the interactive lab
✓Model research → execution → inference optimization

Automatic optimization. Your objective.

MiniCPM5-2B · 512 input / 128 output tokens
1,000 prompts · concurrency 128 · seed 2026

Best results vs. model baseline

512 input / 128 output tokens · 1000 prompts · concurrency 128 · seed 2026 · Measured vllm baseline (6q8zo485)

Output throughput0.4%throughput upliftBest 14,361.4 (vLLM) vs. baseline 14,301.4 tok/s
p95 end-to-end latency6.2%latency reductionBest 1,299.2 (vLLM) vs. baseline 1,385.1 ms
p95 time to first token5.9%latency reductionBest 493.9 (vLLM) vs. baseline 525.1 ms
003,87817,7547,75535,50711,63353,26115,51071,014vLLM: 7,323 tokens/s · 2,512 msvLLM: 6,383 tokens/s · 3,588 msvLLM: 1,929 tokens/s · 8,391 msvLLM: 1,223 tokens/s · 13,315 msvLLM: 5,676 tokens/s · 3,653 msvLLM: 7,420 tokens/s · 2,717 msvLLM: 3,807 tokens/s · 5,511 msvLLM: 6,748 tokens/s · 2,697 msvLLM: 5,218 tokens/s · 3,979 msvLLM: 10,630 tokens/s · 1,582 msvLLM: 10,866 tokens/s · 1,761 msvLLM: 11,408 tokens/s · 1,708 msvLLM: 8,271 tokens/s · 2,330 msvLLM: 7,621 tokens/s · 2,548 msvLLM: 13,080 tokens/s · 1,717 msvLLM: 6,747 tokens/s · 2,427 msvLLM: 12,346 tokens/s · 1,727 msvLLM: 10,840 tokens/s · 1,752 msvLLM: 8,734 tokens/s · 2,158 msvLLM: 7,635 tokens/s · 2,439 msvLLM: 6,268 tokens/s · 4,052 msvLLM: 1,974 tokens/s · 8,217 msvLLM: 1,252 tokens/s · 12,911 msvLLM: 5,733 tokens/s · 3,625 msvLLM: 7,381 tokens/s · 2,915 msvLLM: 3,772 tokens/s · 5,563 msvLLM: 6,782 tokens/s · 2,645 msvLLM: 5,438 tokens/s · 3,810 msvLLM: 10,660 tokens/s · 1,584 msvLLM: 12,148 tokens/s · 1,753 msvLLM: 11,502 tokens/s · 1,736 msvLLM: 11,821 tokens/s · 1,728 msvLLM: 11,762 tokens/s · 1,824 msvLLM: 8,641 tokens/s · 2,137 msvLLM: 13,133 tokens/s · 1,771 msvLLM: 8,959 tokens/s · 2,256 msvLLM: 9,043 tokens/s · 2,222 msvLLM: 9,803 tokens/s · 2,076 msvLLM: 13,103 tokens/s · 1,691 msvLLM: 12,209 tokens/s · 1,720 msvLLM: 12,964 tokens/s · 1,532 msvLLM: 13,045 tokens/s · 1,299 msvLLM: 13,059 tokens/s · 1,810 msvLLM: 7,399 tokens/s · 2,501 msvLLM: 6,341 tokens/s · 3,847 msvLLM: 1,938 tokens/s · 8,329 msvLLM: 1,233 tokens/s · 13,127 msvLLM: 5,625 tokens/s · 3,694 msvLLM: 7,407 tokens/s · 2,810 msvLLM: 3,805 tokens/s · 5,454 msvLLM: 6,802 tokens/s · 2,632 msvLLM: 5,294 tokens/s · 3,918 msvLLM: 10,415 tokens/s · 1,580 msvLLM: 8,853 tokens/s · 2,226 msvLLM: 9,846 tokens/s · 1,869 msvLLM: 10,602 tokens/s · 1,738 msvLLM: 8,438 tokens/s · 2,196 msvLLM: 10,646 tokens/s · 1,614 msvLLM: 7,415 tokens/s · 2,451 msvLLM: 8,177 tokens/s · 2,057 msvLLM: 7,984 tokens/s · 2,653 msvLLM: 8,557 tokens/s · 2,432 msvLLM: 10,868 tokens/s · 1,753 msvLLM: 7,304 tokens/s · 2,541 msvLLM: 6,448 tokens/s · 3,899 msvLLM: 1,938 tokens/s · 8,350 msvLLM: 1,258 tokens/s · 12,911 msvLLM: 5,645 tokens/s · 3,747 msvLLM: 7,408 tokens/s · 2,821 msvLLM: 3,833 tokens/s · 5,491 msvLLM: 6,782 tokens/s · 2,635 msvLLM: 5,242 tokens/s · 3,988 msvLLM: 10,730 tokens/s · 1,562 msvLLM: 11,172 tokens/s · 1,711 msvLLM: 11,325 tokens/s · 1,705 msvLLM: 8,912 tokens/s · 2,566 msvLLM: 8,350 tokens/s · 2,439 msvLLM: 13,092 tokens/s · 1,631 msvLLM: 12,257 tokens/s · 1,700 msvLLM: 12,843 tokens/s · 1,693 msvLLM: 11,965 tokens/s · 1,735 msvLLM: 9,912 tokens/s · 2,216 msvLLM: 11,284 tokens/s · 1,730 msvLLM: 4,589 tokens/s · 3,950 msvLLM: 7,788 tokens/s · 2,423 msvLLM: 5,975 tokens/s · 2,946 msvLLM: 5,599 tokens/s · 3,686 msvLLM: 1,290 tokens/s · 13,961 msvLLM: 1,523 tokens/s · 11,808 msvLLM: 6,766 tokens/s · 2,610 msvLLM: 3,518 tokens/s · 4,569 msvLLM: 12,789 tokens/s · 1,745 msvLLM: 12,287 tokens/s · 1,733 msvLLM: 7,488 tokens/s · 2,438 msvLLM: 12,893 tokens/s · 1,447 msvLLM: 10,342 tokens/s · 2,381 msvLLM: 14,361 tokens/s · 1,531 msvLLM: 14,094 tokens/s · 1,361 msTensorRT-LLM: 9,643 tokens/s · 1,964 msTensorRT-LLM: 11,388 tokens/s · 2,103 msTensorRT-LLM: 11,304 tokens/s · 1,838 msTensorRT-LLM: 10,341 tokens/s · 1,958 msTensorRT-LLM: 6,869 tokens/s · 2,905 msTensorRT-LLM: 3,096 tokens/s · 7,929 msTensorRT-LLM: 10,238 tokens/s · 1,653 msTensorRT-LLM: 1,945 tokens/s · 8,382 msTensorRT-LLM: 9,768 tokens/s · 2,124 msTensorRT-LLM: 11,294 tokens/s · 1,911 msTensorRT-LLM: 8,915 tokens/s · 1,904 msTensorRT-LLM: 8,917 tokens/s · 1,894 msTensorRT-LLM: 10,297 tokens/s · 1,963 msTensorRT-LLM: 9,995 tokens/s · 1,929 msTensorRT-LLM: 10,550 tokens/s · 2,094 msTensorRT-LLM: 9,980 tokens/s · 1,988 msTensorRT-LLM: 10,304 tokens/s · 1,959 msTensorRT-LLM: 6,726 tokens/s · 2,921 msTensorRT-LLM: 2,964 tokens/s · 8,470 msTensorRT-LLM: 10,346 tokens/s · 2,014 msTensorRT-LLM: 1,914 tokens/s · 8,545 msTensorRT-LLM: 9,393 tokens/s · 2,191 msTensorRT-LLM: 10,476 tokens/s · 2,033 msTensorRT-LLM: 10,738 tokens/s · 1,976 msTensorRT-LLM: 9,201 tokens/s · 1,833 msTensorRT-LLM: 9,813 tokens/s · 1,969 msTensorRT-LLM: 9,783 tokens/s · 2,206 msTensorRT-LLM: 9,894 tokens/s · 2,129 msTensorRT-LLM: 10,175 tokens/s · 1,915 msTensorRT-LLM: 6,735 tokens/s · 2,951 msTensorRT-LLM: 2,601 tokens/s · 9,446 msTensorRT-LLM: 10,391 tokens/s · 1,913 msTensorRT-LLM: 1,895 tokens/s · 8,697 msTensorRT-LLM: 9,512 tokens/s · 2,170 msTensorRT-LLM: 11,160 tokens/s · 1,960 msTensorRT-LLM: 8,626 tokens/s · 2,159 msTensorRT-LLM: 10,246 tokens/s · 1,971 msTensorRT-LLM: 10,018 tokens/s · 1,919 msTensorRT-LLM: 10,470 tokens/s · 2,047 msTensorRT-LLM: 11,131 tokens/s · 2,011 msTensorRT-LLM: 10,092 tokens/s · 2,015 msTensorRT-LLM: 6,772 tokens/s · 2,939 msTensorRT-LLM: 2,795 tokens/s · 9,006 msTensorRT-LLM: 10,685 tokens/s · 1,939 msTensorRT-LLM: 1,941 tokens/s · 8,417 msTensorRT-LLM: 9,479 tokens/s · 2,154 msTensorRT-LLM: 9,609 tokens/s · 2,221 msTensorRT-LLM: 10,453 tokens/s · 2,023 msTensorRT-LLM: 10,915 tokens/s · 1,874 msTensorRT-LLM: 10,167 tokens/s · 2,010 msTensorRT-LLM: 9,958 tokens/s · 1,943 msTensorRT-LLM: 10,893 tokens/s · 2,068 msTensorRT-LLM: 11,062 tokens/s · 2,056 msTensorRT-LLM: 10,015 tokens/s · 2,026 msTensorRT-LLM: 6,628 tokens/s · 2,951 msTensorRT-LLM: 2,767 tokens/s · 8,970 msTensorRT-LLM: 10,054 tokens/s · 2,050 msTensorRT-LLM: 1,902 tokens/s · 8,563 msTensorRT-LLM: 9,414 tokens/s · 2,244 msTensorRT-LLM: 10,672 tokens/s · 1,896 msTensorRT-LLM: 251 tokens/s · 65,754 msTensorRT-LLM: 9,474 tokens/s · 2,362 msTensorRT-LLM: 3,949 tokens/s · 5,015 msTensorRT-LLM: 9,924 tokens/s · 1,800 msTensorRT-LLM: 1,909 tokens/s · 8,536 msTensorRT-LLM: 9,647 tokens/s · 1,960 msTensorRT-LLM: 7,607 tokens/s · 4,719 msTensorRT-LLM: 10,409 tokens/s · 2,200 msTensorRT-LLM: 1,393 tokens/s · 11,776 msTensorRT-LLM: 9,998 tokens/s · 2,003 msTensorRT-LLM: 9,872 tokens/s · 1,962 msTensorRT-LLM: 10,543 tokens/s · 2,115 msTensorRT-LLM: 9,881 tokens/s · 2,062 msTensorRT-LLM: 10,007 tokens/s · 2,002 msTensorRT-LLM: 6,660 tokens/s · 2,942 msTensorRT-LLM: 2,758 tokens/s · 9,010 msTensorRT-LLM: 9,881 tokens/s · 1,964 msTensorRT-LLM: 1,900 tokens/s · 8,621 msTensorRT-LLM: 9,402 tokens/s · 2,258 msTensorRT-LLM: 10,193 tokens/s · 2,099 msp95 E2E latency · msOutput throughput · tokens/s
178 trials · hover or tap a point to inspect
178 / 178
Recorded trials · data context

Hardware, revision, and runtime metadata are incomplete. Inspect context before comparing.

Explore benchmarks

Connected by design.

NemotronSGLangvLLMTensorRT-LLMNebiusServerless AILLM OptimizerDynamo

A model choice is
only the beginning.

A model’s reputation doesn’t tell you how it will serve your workload.

Loading recorded benchmark metrics…

Qwen3.5-4B · 512 in / 128 out · c128
Workload and source context

896 prompts · seed 2026. Recorded trials; hardware, revision, and runtime metadata are incomplete.

One connected workflow.
Every decision is traceable.

From a model candidate to a configuration you can inspect.
Explore how the pieces work together.

Architecture walkthrough

Select components for details

User experience

Research with context.

Nemotron-assisted assessment brings model metadata and sourced research into the selection process.

Inspect the research workflow

Compare the tradeoffs.

Fixed workloads, explicit objectives, and Pareto analysis make serving configurations easier to assess.

Explore throughput and latency

A better configuration
starts with your workload.

Keep the benchmark fixed. Explore serving parameters. Compare throughput and latency with the context that makes the measurements meaningful.

  • Fixed input lengths, concurrency, and seed
  • Framework-specific parameter search
  • Explicit objectives and Pareto analysis
See where optimization fits

Preparation profiles cover vLLM, SGLang, and TensorRT-LLM. Model and runtime compatibility must be checked.

Serving configuration
Workload stays fixedModel revision · request concurrency · tokens
FIXED
SEARCH DIMENSIONWHAT IT CONTROLS
01max_num_seqsServing capacity
02max_num_batched_tokensToken budget
03gpu_memory_utilizationMemory allocation
MEASUREThroughput Latency
Parameter names are illustrative; domains come from the prepared request.

Watch a search
learn where to look.

Step through actual TPE, Gaussian-process, and random sampler traces on the same synthetic objective. Change the seed. Inspect every observation.

Open the search playground

Offline simulations. Real algorithms. No GPU inference.

Explore observe refine Open lab

Measurements need context.

We publish inference comparisons only when model revision, hardware, runtime, and workload can be traced. Explore recorded optimization runs with their available metadata and separately labeled simulations.

Explore the benchmark interface Read the publication criteria

Bring your workload into focus.

Define the comparison before searching for a configuration.

Plan your evaluation