Llama-3.3-70B on Jetson Thor: frontier-class onboard inference

Meta's Llama-3.3-70B is a 70-billion-parameter, text-only, multilingual model with a 128K context window. Quantized to 4-bit it's about 40 GB — small enough to load onto a single NVIDIA Jetson Thor (T5000), which carries 128 GB of LPDDR5X. That makes genuinely large-model inference possible onboard a robot or vehicle, with no cloud round-trip. Here are the real numbers, including the ones that constrain throughput.

tldr

Llama-3.3-70B (Llama 3.3 Community License, 128K context, text-only) is Hardened in our registry. A 4-bit quant (~40 GB) fits comfortably in Jetson Thor's 128 GB. But single-stream decode is memory-bandwidth-bound: at Thor's 273 GB/s, reading a ~40 GB weight set once per token caps you near ~7 tokens/sec before any speculative-decoding or batching tricks. The win here is capability onboard, not raw speed. Verify the model on its live artifact page.

parameters
70 B (dense)
context
128K tok
modality
text-only, multilingual
license
Llama 3.3 Community
on-disk
~140 GB BF16 · ~40 GB 4-bit
target
Jetson Thor T5000 (128 GB)
tier (registry)
HARDENED
released
2024-12 (Llama 3.3)

Does it fit? Yes.

Llama-3.3-70B at BF16 is about 140 GB; a 4-bit quantization (e.g. Q4_K_M / AWQ / NVFP4) lands around 35–43 GB. NVIDIA's Jetson Thor module — the T5000, part of the AGX Thor Developer Kit — ships with 128 GB of LPDDR5X and a Blackwell GPU with FP4 support (up to ~2070 sparse FP4 TFLOPS). A ~40 GB weight set leaves ample headroom for KV cache and the rest of the system. So the model fits — that part is real and unremarkable.

How fast? Bandwidth decides.

The honest constraint is memory bandwidth, not compute. Autoregressive decoding reads (most of) the weights once per generated token, so single-stream throughput is bounded by memory bandwidth ÷ active weight bytes. Thor's memory bandwidth is 273 GB/s. Against a ~40 GB 4-bit weight set that's a ceiling of roughly 273 / 40 ≈ 7 tokens/sec before overheads — and real-world will be a bit under that. You can push higher with speculative decoding, smaller draft models, MoE sparsity (not applicable to a dense 70B), or batching multiple requests — but for a single interactive stream, single-digit tok/s is the regime. Any "tens of tokens/sec for a dense 70B on a 273 GB/s device, single stream" claim should be treated skeptically; the bandwidth math doesn't support it.

That's still useful. For onboard reasoning where the alternative is a cloud round-trip (or no model at all), a few tokens per second of a real 70B — air-gapped, on the vehicle — is the capability that matters.

Specs & sources

more releases