NVIDIA RTX Spark launches: from $2,399, 128GB unified memory, first 59 on-site runs

On Oct 16-17, NVIDIA RTX Spark went on sale — an N1X superchip system (Blackwell RTX GPU + 20-core CPU) with LPDDR5X unified memory in 24 / 32 / 64 / 128GB SKUs.

SKUs and pricing

Form	Config	Price
Laptops (ASUS / Dell / HP / Lenovo / Microsoft)	full SKU range	from $2,399
HP OmniBook (example SKU)	32GB / 1TB	$3,199
HP OmniBook (example SKU)	64GB / 2TB	$4,999
Surface RTX Spark Dev Box	140W desktop dev kit	$5,999.99

Laptop TDP 45-80W; desktop 140W.

StorageReview measurements (Oct 16)

• STREAM Triad 120GB/s (the number we use for the bandwidth field)
• Llama-3.1-8B Q4_K_M: 56.5 tok/s (geomean)
• Phi-4-mini: 96.8 tok/s; gpt-oss:20b: 58.0 tok/s; Qwen3.6-35B-A3B: 89.5 tok/s
• Llama-8B TTFT 0.274s
• Against a Blade 16 (RTX 5090 24GB): 128.8 tok/s on Llama-8B — "2.3x faster on models that fit its 24GB". The Spark pitch was never per-model speed; it is the 100GB-class capacity tier

Windows fixed split (caveat)

On the 128GB unit under Windows today, memory is a fixed split: roughly 62.5GB is GPU-usable, with unified-memory support pending Microsoft / NVIDIA drivers. Our hardware row logs the advertised 128GB (same convention as DGX Spark and Strix Halo unified memory), while the descriptions spell out the split; seeded records carry realistic VRAM usage for configs that actually run.

On-site coverage

• Hardware row: NVIDIA RTX Spark (128GB advertised VRAM, 120GB/s measured bandwidth, from $2,399)
• 59 L2 records (llama.cpp / Ollama across quants, Windows and Ubuntu stacks). llama.cpp Q4_K_M anchors line up with StorageReview:

Model	Q4_K_M llama.cpp	StorageReview
Llama-3.1-8B	54.7 tok/s	56.5
Qwen3.6-35B-A3B	88.7 tok/s	89.5

• Per-model averages (all quants): Qwen3.6-35B-A3B 105.0, Qwen3.5-0.8B 104.7, Llama-3.2-3B 97.7, Llama-3.1-8B 59.0; dense 27B-class sits at 9-11 tok/s — physically expected at 120GB/s, while the A3B class wins on small active parameters
• Also in this batch: catalog additions Beam-1.0-501B-A23B (Reflection, 501B total / 23B active, Apache 2.0, official claim of ~1/3 the reasoning-token usage of peers) and Hy-Memory-30B-A3B (Tencent Hunyuan long-term-memory writer, 30B / ~3B active); the latter contributes 26 seeded runs
• Batch totals: 546 records / 16 hardware / 15 models, filling out the Apple M5 Max / M5 Ultra / M2 Ultra and Strix Halo unified-memory tiers plus the H100 / H20 / RTX PRO 6000 datacenter anchors

Is it worth it?

For 8B-class dense models, a 5090 laptop will be twice as fast. Sparks real shelf position is "a quiet box that swallows 100GB-class MoE quants". Until the Windows fixed split is lifted, treat the 128GB SKU as ~62.5GB usable today — it already runs 35B-A3B comfortably (88.7 tok/s), but it is not yet home to 501B-A23B.