DGX Spark vs RTX PRO 6000 vs Mac Studio vs Strix Halo for local LLMs

Quick answer

Which is fastest: DGX Spark, RTX PRO 6000, Mac Studio or Strix Halo?
On the overlapping 2026 models in RigBase, RTX PRO 6000 leads when the model fits its 96 GB — e.g. Qwen3.8 27B at 150 tok/s. DGX Spark, Mac Studio and Strix Halo trade some speed for 128–512 GB unified memory and lower whole-box power. Sources use different engines and quants, so compare orders of magnitude.

Last updated: 28 Oct 2026

Four common “big memory” options for home and lab local LLMs. Every cell is a measured RigBase result; engines and quants differ by source, so use the table for ballpark speed, not single-digit rankings.

Quick answer: On Qwen3.8 27B, an RTX PRO 6000 measured 150 tok/s vs 14.5 on a DGX Spark (different quants/engines — see table). On gpt-oss 20B, RTX PRO 6000 reached 286.91 tok/s; a Mac Studio measured 116.08. Pick PRO 6000 for raw CUDA speed at 96 GB; Spark / Mac / Strix when you need 128 GB+ unified memory in a small box.

Measured tok/s on current models

Model	DGX Spark	RTX PRO 6000	Mac Studio	Strix Halo
gpt-oss 120B	44.4 source	196.31 MXFP4 source	79.68 MXFP4 source	55.8 MXFP4 source
Qwen3.8 27B	14.5 general, short prompt source	150 ⚡ NVFP4 (W4A16) + MTP (5 draft source	50.89 BF16, 8,192-token context source	12.6 Q4_K_M source
gpt-oss 20B	96.2 source	286.91 MXFP4 source	116.08 MXFP4 source	–
Qwen3.8 Flash-Next	54.4 NVFP4 source	76.337 IQ3_S, 32K prompt, all on th source	–	82 custom int4 experts / 6-bit source
GLM-5.3 Flash	105.3 NVFP4 source	265.9 ⚡ EXL3 4-bit + DFlash2 source	–	–
DeepSeek V4.1 Flash	31.6 EXL3 2.9-bit source	–	30.8 ⚡ not stated source	–
Qwen3.6 35B-A3B	89.5 Ollama default, 64K context, source	–	–	62 Q4_K_M (Unsloth) source
Gemma 4 26B-A4B	30.8 no speculative decoding source	–	–	51.8 4-bit source

⚡ = speculative decoding (or the only available result used it). Blank = no measured result for that pair yet. Same-method power comparison (Spark vs Mac vs Strix only): tok/s per watt page.

How to choose

RTX PRO 6000 (96 GB): fastest CUDA single-card option in this set when the model fits.
DGX Spark (128 GB GB10): small CUDA box, multi-Spark recipes for huge MoE; see all Spark recipes.
Mac Studio (up to 512 GB): most unified memory; MLX/llama.cpp; strong on large dense models.
Strix Halo (128 GB): cheapest path to 128 GB unified; prefer Vulkan; pair with MTP on dense 27B.

More: all comparisons · Mac vs PC · VRAM calculator

FAQ

Which is fastest: DGX Spark, RTX PRO 6000, Mac Studio or Strix Halo?

On the overlapping 2026 models in RigBase, RTX PRO 6000 leads when the model fits its 96 GB — e.g. Qwen3.8 27B at 150 tok/s. DGX Spark, Mac Studio and Strix Halo trade some speed for 128–512 GB unified memory and lower whole-box power. Sources use different engines and quants, so compare orders of magnitude.

When do I pick the DGX Spark over an RTX PRO 6000?

When you need more than ~96 GB for one model (or multi-Spark recipes for 300B+ MoE), CUDA/vLLM, and a small workstation. See the DGX Spark recipes page for 1–4 Spark measured setups.

Where does Strix Halo fit?

128 GB unified memory at PC prices; dense 27B decode is single-digit without MTP. Close to a Spark on some MoE models, slower on dense ones. See the AMD guide.

Not enough VRAM? Rent before you buy. Run a 120B model on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.