THE SAME WORKLOAD. DIFFERENT POSSIBILITIES.

See how your
setup stacks up.

Put real measurements side by side. Explore generation speed, response time and the hardware behind every run.

REGISTRY RUNS5Included in this website buildRun a benchmark

Your file is read in this browser tab only. It is never uploaded or added to community rankings, and is cleared when you reload.

Or paste a packaged result

All differences are measured against this run.

Everything matches: the same model, weights, quantization, workload, runtime, placement and transport. Only these can be ranked against each other.

Ranking is stricter than looking. The leaderboard below only ever ranks runs of the same declared workload, whatever is chosen here.

MATCHING WORKLOAD

deepseek-ai/DeepSeek-R1-Distill-Qwen-14B

Q4_K_M · chat-small-v2

What makes a comparison compatible?

Model, weights source, immutable revision, quantization, profile ID and checksum, and GuideLLM version must match. So must the transport, the serving arrangement, whether the weights were split across CPU and GPU, and the tokenizer that counted the tokens: a run measured through the Windows–WSL bridge carries that bridge's latency inside every number, and token counts are the counts of whichever tokenizer produced them. Hardware, operating systems and runtime settings may differ. This compares complete setups, not the effect of one isolated component, and token rates across different models do not describe the same semantic work.

Weights source
Canonical model repository (legacy record)
Weights revision
sha256-manifest:39e40f3bbefbfb4786d38b9a59d5beff6043eaf0ee92a2539bedea5fe724e6d5
Profile checksum
sha256:efba7d0eff0bc8c639f94af35de03e6d6f5869ea54ec6749f58808d6fd02e4ca
GuideLLM
0.7.3
Transport
Windows–WSL bridge · its latency is inside every number here
Tokenizer
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B · loaded before the run

02 / BUILD YOUR COMPARISON

Your reference is ready.

0 compatible runs

No other compatible runs are available in this build. Your measurements are shown below. Community comparisons appear after matching evidence is submitted, reviewed and included in a website build.

4 runs excluded from direct comparison
  • 1× AMD Radeon RX 9070 · 16 GB eachDifferent model · Different weights revision · Different or unrecorded tokenizer source
  • 1× AMD Radeon RX 9070 · 16 GB eachDifferent model · Different weights revision · Different or unrecorded tokenizer source
  • 1× AMD Radeon RX 9070 · 16 GB eachDifferent model · Different weights revision · Different or unrecorded tokenizer source
  • 1× AMD Radeon RX 9070 · 16 GB eachDifferent weights revision · Different context length
REFERENCE RUN

1× AMD Radeon RX 9070 · 15.92 GB each

GPUlmstudio LM Studio 0.4.28 (build 2)BRIDGED
52.99output tok/s
CPU
AMD Ryzen 7 5800X3D 8-Core Processor8 cores · 16 threads
Inference GPU
AMD Radeon RX 9070GPU only
System RAM
32 GBinstalled
VRAM
15.92 GB each
Measured requests
50 / 50 successful
Error rate
0%
Setup details
Serving CPU
AMD Ryzen 7 5800X3D 8-Core Processor
Serving OS
Microsoft Windows 11 Pro Insider Preview 10.0.26220
Accelerator
1 × AMD Radeon RX 9070, 15.92 GB each
Quantization
Q4_K_M
Context length
8192 tokens
Serving runtime
llama.cpp LM Studio 0.4.28 (build 2) · hosted by LM Studio 0.4.23
Weight placement
GPU only
Serving arrangement
Windows host served, WSL2 runner measured
Transport
Windows–WSL bridge · its latency is inside every number here
Tokenizer
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B · loaded before the run
Served identifier
deepseek-r1-distill-qwen-14b
Runtime arguments
None recorded
Run ID
run-0d15220268ae

HOW RESPONSIVE DOES IT FEEL?

A closer look at latency.

p50 is the median. p95 describes the slower end of observed latencies.

RESPONSE TIME

First token, p50

↓ Lower is better
  1. 1× AMD Radeon RX 9070 · 15.92 GB eachReference643.7 ms

    Reference measurement

Final measured-window latency. Each chart has its own zero-based scale.

STREAMING SMOOTHNESS

Between tokens, p50

↓ Lower is better
  1. 1× AMD Radeon RX 9070 · 15.92 GB eachReference16.4 ms

    Reference measurement

Final measured-window latency. Each chart has its own zero-based scale.

SAME DECLARED WORKLOAD

Community run leaderboard

1 matching runs

Ranks apply only to the matching runs and hardware filters below. Repeat runs count separately, this is not a ranking of unique machines or people. Private files are excluded.

Output throughput · Higher is better · chat-small-v2
RankSetup / submitterOutput throughputSuccessful / attemptedError rateRecord
01AMD Ryzen 7 5800X3D 8-Core Processor · AMD Radeon RX 9070RAM 32 GB · VRAM 15.92 GB each · deepseek-ai/DeepSeek-R1-Distill-Qwen-14B · Q4_K_M · LM Studio 0.4.28 (build 2) · XatiivReference52.99 tok/s50 / 500%Evidence ↗

MAKE THE NEXT COMPARISON POSSIBLE

Add evidence. Help everyone choose.

Run a fixed profile, review your bundle and submit it through the local panel. Community results appear after review, merge and a website build.

Run your benchmark ↗