This review is built from NVIDIA’s published specifications and pricing, plus third-party testing from ServeTheHome, HotHardware, JetsonHacks and others, rather than hands-on use. Figures are attributed in the text.
The NVIDIA Jetson AGX Thor developer kit is a robotics computer with 128GB of memory, a Blackwell GPU and a power range of 40 to 130W. It went on general sale in August 2025 at $3,499, and in 2026 it became the reference platform for humanoid robots and other machines that need to run large models on board.
Here is the honest verdict for this site’s readers. The Jetson AGX Thor is excellent at what it was built for: running vision, language and action models on a robot, with real-time I/O and a power budget a battery can sustain. As a desktop box for local chat models, it is the wrong purchase. NVIDIA’s own DGX Spark has the same 273 GB/s memory bandwidth, a faster CPU and better networking, and is quicker at single-user LLM inference in every published side-by-side test. Buy Thor if your model has wheels or legs.
Table of Contents
- What the Jetson AGX Thor Offers
- Jetson AGX Thor Price and Availability
- Jetson AGX Thor LLM Benchmarks
- Jetson AGX Thor Software Got Faster After Launch
- Jetson AGX Thor vs DGX Spark
- Jetson AGX Thor vs Jetson AGX Orin
- Why the TOPS Figure Misleads
- What the Jetson AGX Thor Is Really For
- Jetson AGX Thor Ports and Sensors
- Jetson AGX Thor Power Modes
- Developer Experience and Ecosystem
- Models That Suit 128GB at 273 GB/s
- The Jetson AGX Thor as a Desktop LLM Box
- Who Should Buy What
- FAQ
- Setting Up the Jetson AGX Thor
- Verdict
- Sources and Further Reading

What the Jetson AGX Thor Offers
Thor is the successor to the Jetson AGX Orin, NVIDIA’s long-running robotics module. The developer kit pairs the T5000 module with a carrier board, storage and cooling.
| Specification | Jetson AGX Thor Developer Kit (T5000) |
|---|---|
| GPU | Blackwell, 2,560 CUDA cores |
| Memory | 128GB LPDDR5X |
| Memory bandwidth | 273 GB/s on a 256-bit bus |
| Power range | 40 to 130W, configurable |
| Storage | 1TB NVMe |
| Networking | QSFP port for high-speed Ethernet, plus a 5GbE port |
| Price | $3,499 |
TechRadar summarised the kit as 128GB of LPDDR5X with over 270 GB/s of bandwidth, 1TB of storage, HDMI 2.1, Wi-Fi 6E and a 100GbE-capable port. ServeTheHome praised NVIDIA for specifying a wide power band, which matters for robots whose batteries and thermal limits vary.
The memory is the same capacity and bandwidth as NVIDIA’s DGX Spark, which shapes everything that follows about LLM performance.
Jetson AGX Thor Price and Availability
The developer kit lists at $3,499. HotHardware pointed out the jump from its predecessor: the Jetson AGX Orin developer kit cost $1,999 when it launched in 2023. For robotics teams that is a modest increase for the performance gained. For hobbyists it is a large one.
Third-party listings vary. OpenClaw’s advice is sensible: buy as close to the $3,499 list price as you can, and treat listings above $5,000 with suspicion. NVIDIA distributors and robotics specialists such as Seeed Studio carry the kit.
Production modules for robots are sold separately from the developer kit. If you are building a product, the kit is where development starts, not what ships in the final machine.
Jetson AGX Thor LLM Benchmarks
Several outlets have published language model results. They vary with software version, which turns out to matter a great deal on this platform.
| Model | Result | Source |
|---|---|---|
| Llama 3.1 8B | 149.1 tok/s, against NVIDIA’s expected 150.8 | ServeTheHome |
| Qwen 3 32B | Nearly 80 tok/s | HotHardware |
| Llama 3.3 70B, launch software | About 10 to 12 tok/s | HotHardware, Hackster via ThinkRobotics |
| Llama 3.3 70B, later NVIDIA software | 88.62 tok/s with speculative decoding | NVIDIA via HyperAI |
ServeTheHome’s result is notable mainly because it matched NVIDIA’s claim almost exactly. Several reviewers described the vendor’s numbers as honest for once. The faster figures, such as nearly 80 tokens per second on Qwen 3 32B, come from NVIDIA’s optimised serving stack with batching and low-precision formats, not from a simple llama.cpp run.
For one person typing into a chat window with standard tools, expect speeds closer to what the 273 GB/s bandwidth allows: fast on small models and mixture-of-experts models, slow on dense 70B models.
Jetson AGX Thor Software Got Faster After Launch
The most interesting story about the Jetson AGX Thor is how much faster it became after launch. JetsonHacks explained in December 2025 how the kit had become about 3.5 times faster than at release. NVIDIA later reported a sevenfold improvement in generative AI performance on Llama 3.3 70B through FP4 quantisation and speculative decoding, from launch-day figures around 12 tokens per second to 88.62.
The techniques are worth understanding because they apply elsewhere. Quantising weights to FP4 cuts the bytes read per token. Speculative decoding uses a tiny draft model to guess several tokens ahead and a large model to check them all at once, so when the guesses are right you get several tokens for the cost of one. NVIDIA’s Jetson AI Lab tutorials walk through both using vLLM.
The lesson is that benchmark numbers on Jetson age quickly. Check which JetPack and vLLM versions a figure came from before you compare it with anything.
Jetson AGX Thor vs DGX Spark
This is the comparison most readers of a local AI site need. Both machines have 128GB of memory at 273 GB/s on a 256-bit bus, which OpenClaw notes means they share the same ceiling for token generation, whatever the TOPS figures suggest.
In practice, Spark is ahead. JetsonHacks ran identical llama.cpp tests on both and found DGX Spark averaged 1.36 times the token-generation speed of Thor, with a range of about 1.2 to 1.42 times. ComputingForGeeks reached the same conclusion: Spark is faster at LLM inference on every single-stream test run on both, and has the bigger CPU and network.
| Question | Jetson AGX Thor | DGX Spark |
|---|---|---|
| Memory and bandwidth | 128GB, 273 GB/s | 128GB, 273 GB/s |
| Single-user LLM speed | Slower | About 1.2 to 1.4 times faster |
| Power envelope | 40 to 130W, robot-friendly | Desktop |
| Sensor I/O and real-time kernel | Yes | No |
| Best for | Robots and edge machines | Desk-side development |
ComputingForGeeks summed it up in one line: buy Thor if your model rides on a machine that moves. Our DGX Spark review covers the desktop option in detail.
Jetson AGX Thor vs Jetson AGX Orin
For existing Jetson users, the question is whether to move up from Orin. HotHardware offered a helpful comparison: Orin felt closer to an RTX 3050 and Thor closer to an RTX 5070. That is roughly the performance gap you pay $1,500 more for.
The bigger differences are memory and modern model support. Thor’s 128GB and Blackwell features, including FP4, let it run vision-language-action models and 30B-class language models on board, which Orin cannot do comfortably. ThinkRobotics reported that NVIDIA’s Isaac GR00T humanoid robot models are the main reason buyers at this price choose Thor over Orin.
If your robot runs detection and small models today and will keep doing so, Orin remains capable and much cheaper. If you are building around large multimodal models, Thor is the platform NVIDIA’s robotics stack now targets.
Why the TOPS Figure Misleads
Thor’s headline compute figure is enormous. OpenClaw explained that the quoted 2,070 TFLOPS is a sparse FP4 number, which runs roughly double the dense equivalent and does not predict tokens per second at all.
Language model generation is limited by memory bandwidth, not arithmetic. Each new token requires reading the active model weights from memory, and at 273 GB/s there is a hard ceiling regardless of how many operations the GPU could perform. That is why Thor and Spark, with very different compute figures, generate text at similar speeds.
The compute does matter for other work: vision models, prompt processing, batched serving and the physics and perception workloads robots run alongside language models. Just do not use it to estimate chat speed.
What the Jetson AGX Thor Is Really For
MayhemCode put the use case well: Thor earns its price when you are running something big, on-device, with no cloud fallback, such as a 30B-plus model reasoning about camera input in real time.
A robot cannot always rely on a network connection, and round trips to a cloud server add latency that a moving machine cannot afford. HotHardware noted that very fast token rates are what feed the roughly 10 Hz reasoning loop NVIDIA says robots need to be viable in the real world. Thor is designed to hit that loop on board, inside a power budget a battery can sustain.
That is the market Thor was built for: humanoid robots, autonomous mobile robots, industrial arms with vision, and other machines that must perceive, reason and act locally.
Jetson AGX Thor Ports and Sensors
The developer kit’s connectivity reflects that purpose. ServeTheHome’s teardown found high-speed networking through a QSFP port alongside a Realtek 5GbE NIC, and the board supports the camera and sensor interfaces robots depend on.
ThinkRobotics noted that the included SSD is a WD/SanDisk SN5000S, confirmed by ServeTheHome’s teardown. Storage is ample for models and datasets during development.
For a desktop LLM box, most of this I/O goes unused. For a robot, it is the point of the platform.
Jetson AGX Thor Power Modes
Thor’s configurable 40 to 130W range is what lets the same module power a small mobile robot or a large stationary machine. Lower power modes reduce performance but extend battery life and reduce heat.
ServeTheHome pointed out that if you are building a robot where battery life matters, power consumption is the number to watch. LLM inference at full speed draws heavily, so plan model sizes and duty cycles around your energy budget rather than around peak benchmarks.
Developer Experience and Ecosystem
The software around a robotics platform matters as much as the chip. Here NVIDIA’s advantage is large. Jetson has been the default embedded AI platform for years, so tutorials, forum answers, ROS integrations and container images are plentiful. Hackster’s getting-started guide shows a flashing process that is straightforward for anyone who has used a previous Jetson.
The same CUDA, TensorRT and vLLM tools that run on NVIDIA’s desktop and data-centre hardware run here, which means a model optimised on a workstation can move to the robot with relatively little rework. That continuity is a big part of why robotics teams pick NVIDIA even when cheaper hardware exists.
The flip side is that you are tied to NVIDIA’s release cadence. Performance improvements arrive through JetPack updates, and older JetPack versions miss them. Plan to update regularly during development, then freeze a version before you ship.
Community support for desktop-style tools such as Ollama and llama.cpp exists too, and JetsonHacks used llama.cpp for its comparison with DGX Spark. For the best numbers, though, NVIDIA’s own serving stack is where the optimisations land first.
Models That Suit 128GB at 273 GB/s
The memory pool is large and the bandwidth is moderate. That combination suits mixture-of-experts models and small to mid-sized dense models.
- Vision-language-action models such as Isaac GR00T. The primary workload NVIDIA targets.
- Mixture-of-experts language models. Models like Qwen3.6 35B-A3B read only a few billion parameters per token and run quickly. LLMRun uses it as a reference model for the kit.
- 8B to 32B dense models in FP4 or FP8. Fast enough for real-time reasoning with NVIDIA’s optimised stack.
Dense 70B models fit, but only reach usable speed with speculative decoding and aggressive quantisation through NVIDIA’s serving tools.
The Jetson AGX Thor as a Desktop LLM Box
Some buyers see 128GB of GPU-accessible memory for $3,499 and consider Thor for a home model server. That is usually a mistake. ComputingForGeeks was blunt: if you only want a local chatbot or coding model, neither Thor nor Spark is the cheapest answer, and 128GB Strix Halo boxes cost about half as much per gigabyte.
A Strix Halo mini PC such as the Minisforum MS-S1 Max or GMKtec EVO-X2 runs the same kinds of models at similar speeds, on x86, for less. A DGX Spark is faster if you need CUDA. Thor only makes sense on a desk if that desk is where you develop software that will later run on a robot.
Who Should Buy What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Building a robot or edge machine that runs large models on board | Jetson AGX Thor Developer Kit | Power envelope, sensor I/O, NVIDIA robotics stack | Check price |
| Wanting a CUDA desktop with 128GB | DGX Spark | Faster LLM inference, bigger CPU | Check price |
| Wanting 128GB for local chat at the lowest price | Strix Halo mini PC | Similar speed, about half the cost per GB | Check price |
| Building smaller edge projects | Jetson Orin Nano Super | Far cheaper, same software family | Check price |
Frequently Asked Questions About the Jetson AGX Thor
How much does the Jetson AGX Thor cost?
The developer kit lists at $3,499. Its predecessor, the AGX Orin kit, launched at $1,999.
Is it good for local LLMs?
It runs them well, at 149 tokens per second on Llama 3.1 8B in ServeTheHome’s test. A DGX Spark is faster for desk use, and Strix Halo boxes are cheaper.
How fast is it on 70B models?
About 10 to 12 tokens per second with launch software, and much faster with NVIDIA’s later FP4 and speculative decoding optimisations.
Thor or DGX Spark?
Thor for robots and edge machines. Spark for desk-side development and faster single-user inference.
What memory does it have?
128GB of LPDDR5X with 273 GB/s of bandwidth, the same as DGX Spark.
Why buy it over Orin?
Large multimodal and vision-language-action models, including NVIDIA’s Isaac GR00T, need its memory and Blackwell features.
Setting Up the Jetson AGX Thor
Flash the latest JetPack release using NVIDIA’s installer from a USB drive, as Hackster’s getting-started guide shows. Recent releases carry most of the performance gains reported since launch.
For language models, start with NVIDIA’s Jetson AI Lab containers and vLLM, which include the FP4 and speculative decoding paths. For robotics, install the Isaac stack and the GR00T examples.
Choose a power mode that matches your deployment target early, so benchmarks reflect how the robot will actually run.
Verdict
Buy the NVIDIA Jetson AGX Thor if you are building a robot or edge machine that must run large vision, language or action models on board, within a battery-friendly power budget. For that job it is the most capable platform available, and its software has improved sharply since launch.
The honest caveat is that it is not a good desktop LLM machine for the money. DGX Spark is faster at single-user inference with the same memory, and Strix Halo boxes deliver similar capacity for about half the price per gigabyte.
Next step: if your model will run on something that moves, buy the developer kit near its $3,499 list price and start with NVIDIA’s Jetson AI Lab examples. If it will sit on a desk, look at DGX Spark or a Strix Halo mini PC instead.
Sources and Further Reading
- ServeTheHome: hands-on review, for measured LLM speeds and hardware details.
- HotHardware: Thor tested, for LLM charts and pricing context.
- JetsonHacks: Spark vs Thor llama.cpp benchmarks, for the side-by-side comparison.
- OpenClaw: Thor vs DGX Spark, for the bandwidth and TOPS explanation.
- NVIDIA Jetson AI Lab: vLLM workshop, for FP4 and speculative decoding setup.
On this site: compare the desktop sibling in our DGX Spark review, see a smaller edge option in the Jetson Orin Nano Super review, check cheaper 128GB options in the MS-S1 Max review, and size memory with how much VRAM for AI.




