594B open model runs on 4 DGX Sparks at ~60 tok/s; NVIDIA sells it as a kit
AIFlow_ML · x · 2026-11-17
This Local AI Daily (17 Nov 2026) rounds up measured DGX Spark ecosystem news:
- Inkling 2 Small (594B total / 17B active, Apache-2.0) now runs on four DGX Sparks: two independent 4-box recipes shipped within two days, including Niklas's port of 0xSero's EXL3 engine using a switchless ring topology over the GB interconnect, at roughly 60 tok/s per box. A day later NVIDIA started selling the same four-box ring as a kit.
- Hibana inference engine: a Spark-tuned fast inference engine claiming 2× speedup in 3-bit mode.
- Thinking budget results: adding a "thinking cap" to Qwen3.8-27B lifted GPQA Diamond from 115 to 160–172 depending on hardware (Mac Studio M5 Ultra, DGX Spark, RTX 5090).
Every figure is labeled MEASURED-BY-SOURCE or VENDOR-CLAIM.
More from Infra
- Harvard study: hyperscale data centers used nearly 2% of US electricity in 2024-2025 — Polymarket · 2026-11-17
- Qwen 3.8 27B at 300-500 tok/s on a 5090 laptop, tuned for parallel agents — tebjan · 2026-11-17
- US tightens AI chip export controls to target remote compute access — Dense_Primary629 · 2026-11-17
- Stanford's Azalia Mirhoseini predicts a Cambrian explosion of model-tailored AI chips — agihouse_org · 2026-11-17
- GPU profiling webinar: PyTorch recompile spikes and a claimed 46% throughput gain — cfregly · 2026-11-17
- US data centers now use more power than a midsized country — Dense_Primary629 · 2026-11-17