Sovereign intelligence for all

Research and lessons from our home labs

BuildPrivateAI tests local models on real hardware we own with all of the info you need to accelerate your private AI journey. Because intelligence should be for everyone.

Browse the leaderboard Read the latest

5 signed runs · last measured 32d ago · Qwen3.6-35B-A3B-FP8 on windows-blackwell-workstation

bench@buildprivateai: ~/runs

$ ./bench --model Qwen3.6-35B-A3B-FP8 --seed 42
model Qwen3.6-35B-A3B-FP8@FP8
topology windows-blackwell-workstation · Single GPU
context 4096 tokens
server ready · 192.4 tok/s · peak 63.8 GiB
verdict Great for DGX Sparks view recipe
as-of Jun 30, 2026
> 

For your hardware

What are you running on?

1Vendor

2GPU

3How many

= 24 GB VRAM

or enter VRAM directly GB system RAM (optional) GB

Fits VRAM RAM offload Too big

Saved in your browser only. Never sent to our server.

Benchmarks 5 signed runs

Full leaderboard

Qwen3.6-35B-A3B-FP8 · FP8 Verdict Great for DGX Sparks

windows-blackwell-workstation · Single GPU · 4K ctx

Tok/s 192.4

Peak VRAM 63.8 GiB

Caveat This is a best-of-sweep result, not one continuous server configuration.

as-of Jun 30, 2026 view recipe →

Qwen3.5-122B-A10B-MTP-GGUF · Q4_K_M Verdict Great Mid-Size MoE

windows-blackwell-max-q · Single GPU · 4K ctx

Tok/s 144.5

Peak VRAM 83 GiB

Caveat The 8K shared-prefix bucket needs a service context larger than 8192 because the task suffix raises the full prompt above the shared-prefix length.

as-of Jun 30, 2026 view recipe →

Nex-N2-mini · BF16 Verdict Best Small MoE

windows-blackwell-max-q · Single GPU · 4K ctx

Tok/s 134.5

Peak VRAM 81.5 GiB

Caveat This is a vLLM no-speculation result.

as-of Jun 29, 2026 view recipe →

gemma-4-26B-A4B-it-FP8-Dynamic · FP8 Verdict Great for DGX Sparks

windows-blackwell-max-q · Single GPU · 4K ctx

Tok/s 185.5

Peak VRAM 62.8 GiB

Caveat This is a best-of-sweep result, not one continuous server configuration.