Hussain-Nazary

AI Desktops & Workstations

Whole systems built (or repurposed) for local AI: turnkey AI supercomputers, NPU-powered mini-PCs, and the Mac Studio. All model picks lead with the newest releases (Aug 2026).

NVIDIA DGX Spark

“An AI supercomputer on your desk” — the easiest way to run 100B+ models locally.

Spec	Value
Chip	GB10 Grace Blackwell superchip (20-core Arm CPU: 10× Cortex-X925 + 10× Cortex-A725)
AI compute	1 PFLOPS FP4 (~1,000 AI TOPS), 5th-gen Tensor Cores
Memory	128 GB unified LPDDR5x (256-bit, 273 GB/s) — shared CPU/GPU
Storage	4 TB NVMe SSD
Networking	ConnectX-7 Smart NIC (200 GbE)
Power	~200 W (fanless, desktop form factor)
Price	$4,699 (was $3,999 at announcement, Mar 2025; rose late 2025)
Software	DGX OS (Linux), preloaded NVIDIA AI stack (NIM, CUDA, PyTorch)

Best for: Running up to 200B-parameter models out of the box, prototyping/agent workloads, and teams that want CUDA without buying a server. Mistral Medium 3 and gpt-oss-120B fit fully in memory.

Recommended AI models (newest first):

Flagship fit (128 GB): Mistral Medium 3 (Jul 2026, 48GB-class → fits with room to spare) → ~6–8 tok/s decode (bandwidth-limited) but excellent for agents/batching; gpt-oss-120B Q4 (~70 GB) → ~5–6 tok/s; MiniMax M2.5 (230B, Q4 ~130 GB) is right at the edge — Q3 or API instead; DeepSeek V4 Flash / GLM-5.2 / Kimi K3 are MoE giants → API/enterprise only.
Sweet spot (speed): GLM-4.7-Flash Q4 (30B-A3B, Jan 2026) → ~12–15 tok/s; Gemma 3 27B Q4 (Jul 2026) → ~15 tok/s; Devstral 2 Q4 (24B coding) → ~18 tok/s; Ministral 3 14B → ~25 tok/s.
Multimodal: Gemma 3 27B (native multimodal, Jul 2026); Qwen2.5-VL-72B Q4 (~41 GB) — vision at scale, slow but capable.
Batch/agent workloads: run vLLM with multiple concurrent requests — the 128 GB unified pool shines for multi-user serving of the newest small models.

NVIDIA DGX Station (GB300)

Workstation-grade: a desktop that competes with a rack of GPUs.

Spec	Value
Chip	GB300 Grace Blackwell Ultra superchip (72-core Arm Neoverse V2 CPU + Blackwell Ultra GPU)
AI compute	20 PFLOPS FP4
Memory	748 GB coherent total: 496 GB LPDDR5x (396 GB/s) + 252 GB HBM3e (8 TB/s)
Networking	ConnectX-8 (800 GbE)
Price	~$70,000–110,000 (NVIDIA list ~$69.9K; OEM/reseller pricing higher)

Best for: Teams/researchers running today’s frontier open models (GLM-5.2, Mistral Large 3, MiniMax M2.5, Kimi K3, DeepSeek V4 Flash) fully locally with data privacy, or fine-tuning at workstation scale.

Recommended AI models (newest first):

Frontier open (fits 748 GB): GLM-5.2 Q4 (~744B total → ~420 GB) → ~6–10 tok/s; Mistral Large 3 Q4 (675B total / 41B active, Dec 2025) → ~10–14 tok/s (its 41B active params make it efficient); MiniMax M2.5 Q4 (230B → ~130 GB) → ~15–25 tok/s; DeepSeek V4 Flash and Kimi K3 (open weights, MoE) → Q2/Q3 fit, slow but real.
Workhorse (fast): Gemma 3 27B → 150+ tok/s (HBM path); GLM-4.7-Flash → 180+ tok/s; gpt-oss-120B Q4 → 60–80 tok/s.
Fine-tuning: LoRA/QLoRA on 70B-class fits comfortably in 748 GB.
Embeddings/reranking at scale: Qwen3-Embedding, bge-m3, jina-reranker — huge batch sizes.

AMD Ryzen AI 400 Desktops (Medusa Point)

The first desktop chips officially supporting Microsoft Copilot+ (announced MWC 2026).

Spec	Value
CPU	Zen 5 (Medusa Point), up to 8 cores
NPU	XDNA2, up to 60 TOPS
iGPU	RDNA (up to 16 CU)
Memory	DDR5 (standard desktop DIMMs), 32–128 GB
Price (Aug 2026)	OEM desktops ~$500–1,700 (entry pricing from ~$499; loaded 16 GB/512 GB systems ~$1,660) — sold as complete PCs, no retail CPUs

Best for: Everyday AI in an office/home tower — on-device Copilot+, local transcription, background agents, small models. Not for big LLMs (memory bandwidth is DDR5-class).

Recommended AI models (newest first):

On-NPU (fast, low power): the Qwen3.5 / Gemma 4 sub-12B line (2026), Ministral 3 3B/8B (Dec 2025), Voxtral Transcribe 2 (Feb 2026, on-device STT with diarization); Windows Copilot+ features.
On-GPU/CPU (bigger): Ministral 3 14B Q4 via llama.cpp (iGPU, ~15–25 tok/s depending on RAM speed); Gemma 3 27B Q4 slow but possible on 32 GB+.
Best pairing: add a discrete GPU (e.g., RX 9070 XT) — the NPU handles always-on tasks, the dGPU handles heavy inference.

AI Mini-PCs (Core Ultra / Ryzen AI boxes)

Cheap, quiet, always-on local AI boxes.

Spec	Value
Processors	Intel Core Ultra 200V (Lunar Lake) / 300V (Panther Lake), AMD Ryzen AI 300/400
NPU	48–60 TOPS
Memory	16–96 GB LPDDR5x/DDR5
Form factor	0.5–2 L mini-PC
Price (Aug 2026)	~$400–1,000 mainstream (16–32 GB); $1,000–3,000 loaded (64–128 GB Strix Halo-class); flagship AI boxes: Ryzen AI Halo $3,999, DGX Spark $4,699

Best for: Home servers, privacy-focused assistants, transcription/voice agents, and light RAG. Many folks run Ollama + Open WebUI on these 24/7.

Recommended AI models (newest first):

Comfortable: Ministral 3 3B/8B (Dec 2025) → 20–40 tok/s; Qwen3.5 / Gemma 4 sub-12B (2026) → 15–35 tok/s; Voxtral Transcribe 2 (on-device STT) — the standout voice pick; Gemma 3 12B on 32 GB units.
Stretch: Ministral 3 14B Q4 → 10–20 tok/s (depends on memory config); GLM-4.7-Flash Q4 only on 64 GB+ units.
Embeddings/RAG: Qwen3-Embedding, bge-small, nomic-embed-text.
TTS: Kokoro, Piper — instant.

Apple Mac Studio (M4 Max / M5 Ultra)

See apple-silicon.md for the chips; the Studio is the “big memory” chassis.

Spec	M4 Max Studio	M5 Ultra Studio (reported)
CPU / GPU	16-core / 40-core	36-core / 84-core (reports)
Neural Engine	16-core	32-core (reports)
Unified memory	up to 128 GB	up to 512 GB (768 GB per latest reports)
Memory bandwidth	546 GB/s	~1 TB/s class (reports)
Storage	up to 8 TB	up to 16 TB
Price	from ~$2,000 (base)	from ~$4,000+ (reported)

Best for: The 512 GB M5 Ultra Studio is the only mainstream machine that can hold the newest frontier open models (Mistral Large 3, GLM-5.2, MiniMax M2.5) entirely in memory — the local-inference extreme.

Recommended AI models (newest first):

M5 Ultra (512 GB): MiniMax M2.5 Q4 (230B → ~130 GB) → ~15–20 tok/s; Mistral Large 3 Q4 (675B → ~380 GB) → ~5–8 tok/s; GLM-5.2 Q4 (~744B → ~420 GB) → ~4–6 tok/s — all run locally, which no other desktop can claim; gpt-oss-120B Q4 → ~20–30 tok/s; multiple 70B-class models loaded simultaneously.
M4 Max / M5 Max (128 GB): MiniMax M2.5 Q4 right at the edge (~130 GB) — Q3 or API; Mistral Medium 3 (Jul 2026) → ~10–14 tok/s; gpt-oss-120B Q4 (~70 GB) → ~8–10 tok/s; Gemma 3 27B → ~25–30 tok/s; GLM-4.7-Flash → ~30 tok/s; several 7–30B models at once.
Small/fast: Ministral 3 14B Q4 → 50–70 tok/s; Qwen3.5 / Gemma 4 sub-12B → 80–120 tok/s.
Tools: MLX (Apple’s framework), llama.cpp, Ollama — all well supported.