More
Best rig by budgetRent before you buySubmit a buildNewsletterFree datasetAdvertise
Submit build

Builds / NVIDIA Jetson AGX Thor (Blackwell, unified)

Jetson AGX Thor 128 GB: Qwen3.8 Flash-Next Intel W4A16 AutoRound at 32.1 tok/s on patched vLLM (MTP 3, n-gram table on NVMe)

Last updated:

Source: github.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
Amazon · eBay (used) · Newegg
RAM
128 GB unified
PSU
not stated
Other
Fork of a DGX Spark recipe ported to sm_110: vLLM image with GB10/vLLM patches, Intel W4A16 AutoRound int4 weights + int8 GPTQ lm_head + blockwise-fp8 side layers, fp8 n-gram (PLE) table served from NVMe via mmap; Thor defaults QSA_EXACT_TOPK=1 and Triton GDN decode kernel
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.8 Flash-NextIntel W4A16 AutoRound int4 (+ int8 lm_head, fp8 side layers and PLE table)vLLM (patched fork image), MTP 3, prefix caching on32.1not stated⚡ spec-decode (MTP)context: fresh 1,000-token decode (W1)📏 bench/decode_bench.py, median of 3 (runs 26.9 / 32.1 / 34.7), serve.sh Thor defaults; acceptance 2.37-2.62 tok/step; TTFT 1.38 s cold, ~0.20 s warm. With a pinned ~8K prefix hit + 256 tokens (W2): 40.4 tok/s. An experimental 'UltraFast v16b' branch measured 47.5 tok/s (n=2), not used as the headline

GitHub recipe (Thor port, measured early Oct 2026; repo last updated 7 Oct). Engine is a patched vLLM image, so these are recipe numbers rather than stock vLLM. The same recipe reports about 49 tok/s on a DGX Spark (GB10). Cross-session levels on the Thor swing by several tok/s per the source.

Source: https://github.com/mortacious/qwen3.8-Flash-Thor-AutoRound

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-flash-next-vllm-w4a16-mtp-mortacious.svg)](https://getrigbase.com/builds/jetson-agx-thor-qwen38-flash-next-vllm-w4a16-mtp-mortacious/)

HTML

<a href="https://getrigbase.com/builds/jetson-agx-thor-qwen38-flash-next-vllm-w4a16-mtp-mortacious/" rel="noopener"><img src="https://getrigbase.com/badge/jetson-agx-thor-qwen38-flash-next-vllm-w4a16-mtp-mortacious.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · 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).