Question 1

Is AMD Strix Halo good for local AI?

Accepted Answer

Yes — for the right models. Measured on a GMKtec EVO-X2 AI (Ryzen AI Max+ 395, 128GB): sparse-MoE models run at 48-52 tok/s (GLM 4.7 Flash, Ornith 1.5 35B), small dense models at 30-120 tok/s, but dense 27B+ models crawl (10.7 tok/s) and 70B-class dense manages only 3-5 tok/s. Strix Halo is a MoE and mid-size model platform, not a 70B platform.

Question 2

Strix Halo vs NVIDIA DGX Spark for local LLMs?

Accepted Answer

For dense 70B-class models they are similarly slow (3-5 tok/s vs ~5 tok/s) despite the DGX Spark costing 2x more. For MoE and mid-size models both are usable; Strix Halo wins on price-per-usable-token, DGX Spark on software maturity. Measured numbers in our comparison tables.

Question 3

Why is Strix Halo slower on big dense models than the bandwidth math suggests?

Accepted Answer

Memory-bandwidth arithmetic (256 GB/s / 37GB 70B model = ~7 passes/s) is the ceiling, not the promise — real GPU inference lands at 3-5 tok/s. We published this correction openly after initially overestimating by 5-6x. Bandwidth math gives you the upper bound; architecture decides where you land under it.

Question 4

What VRAM carve should I set on Strix Halo?

Accepted Answer

A 32GB carve (UMA) covers Q4 30B-class models fully in VRAM while leaving system RAM for the OS; larger models use GTT/unified memory automatically. On our 128GB box: 32GB carve runs the 19GB GLM 4.7 Flash at 100% GPU with headroom for a second model.

Question 5

Can Strix Halo run two LLMs at once?

Accepted Answer

Yes — measured: two concurrent GLM 4.7 Flash sessions at 34.3 tok/s each (~69 aggregate, more total throughput than one serial session); three-way splits at 27.5/35.0/39.2. MoE spare capacity is real; use it for parallel agents.

Class	Example (measured)	Speed	Verdict
Sparse MoE 30B-class	GLM 4.7 Flash (30B-A3B)	48.0 tok/s	✅ The sweet spot
Sparse MoE 35B-class	Ornith 1.5 35B-A3B	52.4 tok/s	✅ Fastest big model here
Small dense 7-14B	Ornith 9B, Llama 3.2 3B	31.8-120 tok/s	✅ Instant
Mid dense ~32B	Qwen3.8 27B	10.7 tok/s	⚠️ Usable, patience required
Large dense 70B	Llama 70B Q4	3-5 tok/s	❌ Not a 70B platform

System	Price	Memory	70B speed
Strix Halo (ours)	~$1,999	128GB unified	3-5 t/s
NVIDIA DGX Spark	$3,999	128GB unified	~5 t/s
Mac Studio M4 Ultra	$7,000	192GB	20-25 t/s (est.)
RTX 4090 (offload)	$1,600	24GB + CPU	60-80 t/s (est., hybrid)
H100 80GB	$30,000+	80GB	25 t/s (pure GPU)

Strix Halo for local AI — measured, and corrected where we were wrong

The platform, plainly

What actually runs well (measured, our box)

The correction we publish on purpose

Strix Halo vs the field (70B Q4, third-party + measured)

Why MoE is Strix Halo's killer feature

Configuration notes from our box

Who should buy one