Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16
No-Dragonfly6246 · reddit · 2026-10-27
Embedl's team shipped an optimized Qwen3.8-27B deployment for Jetson AGX Thor: 39.6 tok/s decode and 23.6 tok/s end-to-end at batch 1, with 24.8 GiB peak memory.
Key techniques and results:
- In-house INT4 quantization with roughly half the error of standard AWQ
- FlashHead: the 262k-vocab output layer only scores likely candidates, cutting batch-1 cost
- MTP speculative decoding on top of TensorRT Edge-LLM 0.10.1
- Accuracy vs FP16: GSM8K 96.4 (96.1), IFEval 83.5 (81.9), MMLU-Pro 79.1 (79.5)
Compared to Jetson AI Lab's 35 tok/s llama.cpp Q4KM+MTP setup (no published accuracy), this ships with accuracy, power and per-token energy data. Usable via embedl CLI or an OpenAI-compatible API (1 GB container, 20 GiB weights), plus in-browser testing on a real Thor.
More from Infra
- Anthropic S-1 reveals $800B+ in off-books future compute commitments — GaryMarcus · 2026-10-27
- ~42% of Anthropic's ~11GW contracted compute is NVIDIA GPUs — ns123abc · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27