Models & Benchmarks
21 hosted in the lab, 17 of them loaded in memory right now. 74 more from 6 cloud providers. Every call goes through the same front door.
Cloud providers
Capability matrix
Model counts per capability. Orange runs in the lab. Bone runs in the cloud.
16 in the lab,46 in the cloud
15 in the lab,27 in the cloud
2 in the lab,25 in the cloud
1 in the lab,8 in the cloud
1 in the lab,3 in the cloud
4 in the lab,0 in the cloud
Hearing never leaves the house: 4 / 0.
2 in the lab,3 in the cloud
1 in the lab,1 in the cloud
1 in the lab,0 in the cloud
1 in the lab,3 in the cloud
0 in the lab,31 in the cloud
Tool calling is all cloud: 0 / 31.
Where they live
Every lab model has a physical home: one of three machines in the house.
01 · DGX Spark
The Qwen family on vLLM: Qwen 3.8 27B, Qwen 3.6 35B mixture-of-experts with 3B active, and several 27B variants.
02 · Strix Halo
Whisper base.en for speech. ComfyUI for image and video. Gemma 4 and nomic-embed on Ollama.
03 · Intel NUC
Whisper small, medium and large-v3-turbo. Kokoro text-to-speech. A small Qwen 2.5 3B.
Chat leaderboard
Best measured generation speed for chat models in the lab.
Same prompt, three lanes“How does a call reach a model here?”
DGX Spark
Every app calls one front door. The gateway checks the app’s own key, books the cost to that app and routes the call to a model in the lab or in the cloud. If the gateway goes down, the relay on each machine replays the call straight to the provider, so the answer still arrives.
DGX Spark
Every app calls one front door. The gateway checks the app’s own key, books the cost to that app and routes the call to a model in the lab or in the cloud. If the gateway goes down, the relay on each machine replays the call straight to the provider, so the answer still arrives.
Intel NUC
75 ms to first token
Every app calls one front door. The gateway checks the app’s own key, books the cost to that app and routes the call to a model in the lab or in the cloud. If the gateway goes down, the relay on each machine replays the call straight to the provider, so the answer still arrives.
Mixture-of-experts: 35B parameters, 3B active per token, so it runs fastest of the three.
Lanes stream at each model’s best measured rate. Only the NUC’s 75 ms first token is measured; the Spark lanes’ start delay is illustrative.
Speech, voice, image, video
01 / 04
One 9.9-second clip. Whisper base.en on Strix Halo: 100% word accuracy. Whisper small on the NUC: 95.1%.
02 / 04
Kokoro 82M on the NUC turns text into speech, one sentence at a time as it’s written.
03 / 04
Qwen Image on Strix Halo. 357 images. 0 failures. Median 32 s. Best 19 s.
04 / 04
LTX 2.5 22B on Strix Halo. 55 clips. Median ~4.7 minutes. Best ~3.5 minutes. Upscaled clips: median ~8.4 minutes.
Speech-to-text · one 9.9 s clip · word accuracy
100%
Whisper base.en · Strix Halo
95.1%
Whisper small · Intel NUC
Text-to-speech · Kokoro · Intel NUC
82M
Parameters
Speaks each sentence as soon as it’s written.
Image generation · Qwen Image · Strix Halo
357
Images generated
0
Failures
32s
Median
19s
Best
Video generation · LTX 2.5 22B · Strix Halo
55
Clips rendered
~4.7min
Median per clip
~3.5min
Best clip
~8.4min
Upscaled, median
How it’s measured
Every figure on this page comes from LabScope’s benchmarks. LabScope reruns them on a schedule and keeps every run, so these numbers change as models, drivers and settings change. What you see is the latest run, not a best-case claim.
Read from LabScope, 26 Oct 2026.
Latest run
54tokens/s
Qwen 3.6 35B-A3B · DGX Spark