Chestnut guideUpdated October 23, 2026

Chestnut brings desktop GPU power to comma four.

Chestnut connects an external graphics card to comma four. The ready-to-drive kit uses an AMD Radeon RX 9060 8GB to run openpilot's largest driving models.

See Chestnut at comma See setup

Chestnut is optional. openpilot is still driver assistance. Supervise every drive.

Chestnut ready to drivecomma.ai

Run the larger driving model with Chestnut

comma launched Chestnut on August 12, 2026. The ready-to-drive kit includes Chestnut, the non-XT Radeon RX 9060 8GB, a long USB cable, vehicle power, and mounting strips. comma four remains the official in-car target.

RX 9060GPU in the kit8GB, non-XT model

GPU modeldriving in the carruns on the external graphics card

4Btraining world modelseparate from the driving network

Optionalcompute upgradestandard comma four works without it

Three pieces, one job

comma fourChestnutRX 9060

comma four handles the cameras, driver checks, and the car connection. Chestnut carries road-camera data over USB to the desktop GPU, which prepares the images and runs the larger driving model. The software also has a small-model fallback when the large model is unavailable.

Explore the latest models and training changes

Use the stable Chestnut release

Enter this URL on comma four's Custom Software screen while the car is parked.

Stable Chestnut installerinstaller.comma.ai/commaai/release-chestnutOfficial in-car channel: release-chestnut. Do not paste this into a phone or desktop browser.

See every official Chestnut channel

Assemble and secure it
Install the card in Chestnut. Secure the card and mount the kit in the passenger footwell or under the passenger seat.
Keep airflow clear
The RX 9060 is a desktop card. Do not cover its fans or place loose items against the card.
Connect power and USB
Connect the vehicle power cable. Run the included long USB cable to comma four.
Install the stable release
Enter the URL below on Custom Software. Remove the current software through Settings first if needed.

Chestnut and GPU mounted in the passenger footwell — Passenger-footwell installcomma.ai

Chestnut and GPU mounted under a passenger seat — Under-seat installcomma.ai

What Chestnut adds

The larger Chestnut driving model running on an external GPU
A much larger model-compute budget than the mobile processor provides
A supported path for comma to test and ship larger driving models

What it does not add

Hands-off or unsupervised driving
Automatic support for more cars
A guaranteed improvement on every drive
Official comma 3X or comma three support

Desktop memory moves much more data

Large vision models need compute and fast memory. The RX 9060 has about 9.6 times the Snapdragon 845's peak internal memory bandwidth. It also has dedicated graphics memory and far more GPU compute.

Radeon RX 9060 non-XT compared with Snapdragon 845 MAX in comma four
Specification	comma four	Chestnut kit
Processor	Snapdragon 845 MAX	Radeon RX 9060, non-XT
Memory	LPDDR4X, shared	8GB GDDR6, dedicated
Memory bus	4 × 16-bit = 64-bit	128-bit
Memory speed	Up to 1866 MHz	18 Gbps
Peak bandwidth	About 29.9 GB/s	288 GB/s
GPU	Adreno 630	RDNA 4, 28 compute units
Peak FP32	Not listed by Qualcomm	21.4 TFLOPS
Cooling	Silent MAX cooling	Desktop card; needs open airflow

The 288 GB/s figure is inside the graphics card. It is not the USB cable speed. Peak hardware numbers do not guarantee better driving on every road.

Sources: AMD RX 9060 specifications, Qualcomm Snapdragon 845 brief, and comma four announcement.

The driving model and training world model are different

The driving model interprets camera inputs and plans actions in your car. A world model supplies a simulated environment for training. The 4-billion-parameter figure in comma's October 23 update describes that training world model.

Its architecture table lists 382.905 million parameters for the driving network, while openpilot's release notes still list 880 million. These published counts have not been reconciled; treat them as source-specific figures rather than a verified count of the model installed on your device.

Compare the October 23 architecture table with the current release notes. Parameter count, download size, and runtime memory are different measurements.

More room to learn

comma says it rebuilt its MLSim, data, and training stack so more driving data and training compute can produce better models. Its autonomy team now plans to focus mainly on Chestnut-class models.

More model capacity can support richer visual input and more complex behavior. Earlier openpilot work showed that wider camera input helped with sharp turns and cut-ins. However, comma has not said that a future Chestnut model will use every pixel from every camera.

Read comma's wider-view example

One runtime, more hardware

tinygrad is the model runtime and hardware-access layer used by openpilot. Its current documentation covers Qualcomm and AMD GPUs, including AMD access over USB. That work makes the Chestnut connection practical.

comma's MLSim and data pipeline train the driving models. tinygrad runs model work across the hardware. These parts work together, but they are not the same system.

See tinygrad runtimes See its macOS eGPU release

What does this mean for future models?

comma's launch announcement puts its autonomy work primarily on Chestnut-class models. The latest update shows that work includes both the driving network and the larger training world model. A training-model size is not a promise that the same network will run in your car.

comma four without Chestnut and comma 3X still have their own on-device release path. Those standard installers remain separate from the Chestnut hardware path.

Distillation means teaching a smaller network with a larger one. The sources cited here do not establish that every future on-device model will be distilled, or promise an 8-billion-parameter model by year-end. Our model deep dive separates released changes from those open roadmap questions.

A useful analogy, not proof

comma says comma four plus Chestnut has compute similar to Tesla HW4. Tesla's public support page lists FSD (Supervised) for HW3 and HW4, with some rollouts starting on HW4. That shows why software may differ by hardware. It does not prove a comma distillation plan.

Publicly documented Tesla AI3 and AI4 comparison
Item	AI3 / HW3	AI4 / HW4
FSD (Supervised)	Listed as available	Listed as available
Rollout order	Timing can come later	Some rollouts start here
Memory type, bus, bandwidth	Not published by Tesla	Not published by Tesla

We do not present teardown estimates as Tesla product specifications. Source: Tesla FSD support.

Big Lebowski is one model, not the whole line

A July 1 openpilot change used the name “lebowski model.” The main branch later moved to another large-model object. “Chestnut-class model” is the clearer name for the full line. Chestnut detects the hardware; there is no separate Big Lebowski picker.

Open the Lebowski change See the checked later model pointer

Choose for what works nowNew to comma? Learn the normal comma four experience first. Then check the Chestnut kit and setup before you buy.

Official sources and calculations

Bandwidth calculations: Snapdragon, 3,732 MT/s × 64 bits ÷ 8 = 29.9 GB/s. RX 9060, 18 Gbps × 128 bits ÷ 8 = 288 GB/s. These are peak internal memory-bandwidth calculations, separate from the USB connection and model parameter counts.

The August launch's 30× parameter and 100× compute comparison describes that launch. The October architecture table uses a different model comparison; those multipliers should not be combined into a single current specification.

Introducing Chestnut Chestnut product and setup page Earlier Chestnut release-notes observation Current Chestnut release notes Current Chestnut image and model pipeline October 23 openpilot 0.11.2 model update AMD Radeon RX 9060 Qualcomm Snapdragon 845 brief tinygrad runtime documentation Official Chestnut installer table

Photos: comma.ai. CommaGuide is an independent guide and is not affiliated with comma.ai.

Was this guide helpful?