Builds / NVIDIA Jetson AGX Thor (Blackwell, unified)

Jetson AGX Thor 128 GB (NVIDIA Jetson AI Lab): Qwen3.8 27B Q4_K_M at about 13 tok/s, about 35 with MTP depth 3 (llama.cpp); Nemotron 3.5 Lightning at 115 with DSpark (vLLM)

Last updated: 10 Nov 2026

Source: jetson-ai-lab.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source

GPU(s): NVIDIA Jetson AGX Thor (Blackwell, unified)
VRAM / unified memory (GB): 128
CPU: Jetson T5000 module (Arm)
Amazon · eBay (used) · Newegg
Motherboard: Jetson AGX Thor Developer Kit
Amazon · eBay (used) · Newegg
RAM: 128 GB unified
PSU: not stated
Other: First Jetson board on RigBase. NVIDIA's own Jetson AI Lab tutorial on speculative decoding; llama.cpp llama-server with the GGUF's embedded MTP head; vLLM for DSpark
Reported total cost: not stated
Power: not stated

Measured speed (generation tokens/sec)

Model	Quant	Engine	tok/s	Prompt processing (tok/s)	Details
Qwen3.8 27B	Q4_K_M (unsloth GGUF with MTP weights)	llama.cpp llama-server (CUDA), -ngl all -fa on	13	not stated	📏 NVIDIA's stated baseline before speculation: 'roughly 13 tokens/s' decode (the tutorial gives the rounded figure only)
Qwen3.8 27B	Q4_K_M (unsloth GGUF with MTP weights)	llama.cpp llama-server (CUDA), --spec-type draft-mtp --spec-draft-n-max 3	35	not stated	⚡ spec-decode (MTP)📏 same server with MTP depth 3: 'roughly 13 tokens/s to 35 tokens/s in our testing' (rounded, as published)
Nemotron 3.5 Lightning 30B-A3B	NVFP4 (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4)	vLLM + DSpark drafter (nvidia ...-NVFP4-DSpark), 4 speculative tokens	115	not stated	⚡ spec-decode (DSpark)📏 'In our Jetson testing, Nemotron 3.5 Lightning reached 115 tokens/s with DSpark depth 4'; full Thor command on the Jetson AI Lab model page; no no-speculation baseline given

Vendor-run: NVIDIA's Jetson AI Lab tutorial (page undated; listed as Oct 2026 by edgellm-bench). Figures are the tutorial's rounded numbers; no prompt set, power mode or context is stated, so treat them as indicative. The tutorial's point: on Thor, dense 27B decode is memory-bound, so speculation nearly triples it.

Source: https://www.jetson-ai-lab.com/tutorials/speculative-decoding/

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/jetson-agx-thor-qwen38-27b-mtp-nemotron-lightning-dspark-jetson-ai-lab.svg)](https://getrigbase.com/builds/jetson-agx-thor-qwen38-27b-mtp-nemotron-lightning-dspark-jetson-ai-lab/)

HTML

<a href="https://getrigbase.com/builds/jetson-agx-thor-qwen38-27b-mtp-nemotron-lightning-dspark-jetson-ai-lab/" rel="noopener"><img src="https://getrigbase.com/badge/jetson-agx-thor-qwen38-27b-mtp-nemotron-lightning-dspark-jetson-ai-lab.svg" alt="Measured on RigBase" height="20"/></a>

Preview: · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).