Registry evidence
← Back to results

Registry run record

deepseek-ai/DeepSeek-R1-Distill-Qwen-14B on 1× AMD Radeon RX 9070 · 16 GB each

This record is included in this website build. Inspect its measurements, declared setup and evidence metadata below. Inclusion in a local preview alone does not establish community review.

Compare this setup ↗

Execution device

GPU

Profile

chat-small-v2

Framework

lmstudio LM Studio 0.4.28 (build 2)

Quantization

Q4_K_M

Transport

Windows–WSL bridge

Run started

11/11/2026, 12:54:52 PM

Performance summary

Time to first token, p50
619.52 ms
Time to first token, p95
635.89 ms
Inter-token latency, p50
16.43 ms
Inter-token latency, p95
16.52 ms
Output tokens per second
53.3
Total tokens per second
267.99
Requests per second
0.21
Error rate
0%

Reproducibility record

Model revision
2f3b082b1356a6123f7ed71e65aea340da25d53c
Runtime
lmstudio LM Studio 0.4.28 (build 2)
Runtime arguments
None recorded
GuideLLM
0.7.3
CPU / RAM
AMD Ryzen 7 5800X3D 8-Core Processor · 31.92 GB
Operating system
Microsoft Windows 11 Pro Insider Preview 10.0.26220
Driver version
Not reported
Configuration fingerprint
sha256:2a93c1028c1f0787c6f8e925ca88ebe58f4a5b92019f9ad1f5b76bef90f7de2e
Source commit
Not recorded

What served this model

What this evidence says.

Reading a fact on a machine establishes what that machine reports about itself. None of this independently confirms that the hardware named here did the work, or that the weights were the ones named.

Measured through the Windows–WSL bridge. Its latency is inside every number here and has not been subtracted.

Where this ran

Arrangement
A Windows host served the model; a WSL2 runner measured itstated by the contributor
Serving operating system
Microsoft Windows 11 Pro Insider Preview 10.0.26220stated by the contributor
Serving processor
AMD Ryzen 7 5800X3D 8-Core Processorstated by the contributor
Serving memory
31.92 GBstated by the contributor
Benchmark runner
Linux-6.18.58.1-microsoft-standard-WSL2-x86_64-with-glibc2.43 (wsl2)read on that machineThe machine that ran GuideLLM. On a bridged run this is the WSL2 distribution, and nothing it can see of its host is evidence about what served the model.

What ran the weights

Declared execution device
GPUstated by the contributorWhat the contributor says performed inference. OmoBench does not observe this and cannot confirm it.
Accelerator
1 × amd AMD Radeon RX 9070, 16 GB eachstated by the contributorperformed inference. Finding a device establishes the device, not that inference touched it.
Weight placement
GPU onlystated by the contributor

Which model, and which weights

Canonical model
deepseek-ai/DeepSeek-R1-Distill-Qwen-14Bstated by the contributorThe catalog identity this result is compared under.
Identifier the server answered to
deepseek-r1-distill-qwen-14bstated by the contributorA serving alias. It is not a claim about which weights were loaded.
Catalog status
Cataloguedstated by the contributor
Quantization
Q4_K_Mstated by the contributor
Weights
2f3b082b1356a6123f7ed71e65aea340da25d53cstated by the contributorRecorded as a revision only. Which repository it is a revision of is not separately recorded.

What made it publishable

Eligibility
Not recordednot recordedSchema 1.1 collected none of this. A result from that era was publishable because the model was on the curated list, which was the whole test.

What served it

Framework
lmstudiostated by the contributor
Inference engine
llama.cpp LM Studio 0.4.28 (build 2)stated by the contributorWhat executed the weights.
Hosting application
LM Studio 0.4.23stated by the contributorThe program hosting the engine. Its release version is a different thing from the engine build above.
Startup arguments
Not recordednot recorded

What counted the tokens

Tokenizer source
deepseek-ai/DeepSeek-R1-Distill-Qwen-14Bstated by the contributor
Loaded before the run
Not checked by OmoBench
Tokenizer implementation
Not recordednot recorded
Vocabulary size
Not recordednot recorded
Matches the served weights
Not verifiable from this evidenceA tokenizer that loads is not thereby the one the served weights were trained with, and nothing available to OmoBench can establish that it is.

How the requests travelled

Transport
Through the Windows–WSL bridgeread on that machineBridge windows-stdio-http1-v1.
Protocol
HTTP/1.1, one request per connectionread on that machine
Transport latency
Inside every measurement, never subtractedread on that machineA relay sits inside every request this run made and its cost has not been separated out, so these numbers are comparable with other bridged runs rather than with direct ones.

Evidence bundle

Files retained with this result.

initial run · Xatiiv

result.json

Normalized summary used to render this registry record.

guidellm-benchmarks.json.gz

Raw GuideLLM output

sha256:f227af260bca30ce5e5bd090f91667aabca5429120e907596e022a1d57f3b91c

guidellm.stdout.log

GuideLLM log

sha256:005fa0873912abc18fc41a16a7b3764ef504d7bc171d89b1002eaddebde7a239

environment.json

Environment snapshot

sha256:cb4b1d684e7b6f2a3260e673d46ee541c499c897abb92a487724b4d1be9acfc4

reproduce.md

Reproduction notes

sha256:65d10d32d7fff56e5a7e9a210238a6994b76aa4ba6f808a61c138de101730962