# Hardware Compatibility

> Which GPUs and Macs can serve which bundles, and how the node decides.

Section: Providing inference. Web page: https://docs.katara.com/providers/hardware/

A bundle states the hardware it needs: a video memory floor and a compute capability for NVIDIA, a chip list and a unified memory floor for Apple Silicon. Your machine either meets a profile or it does not, and the node works that out on the machine itself before it advertises anything.

## How Compatibility Is Decided

Compatibility checks the operating system, supported runtime, GPU architecture or Apple chip, and the profile’s memory requirement. Memory budgets must include the model, context cache, runtime working memory and headroom. Memory does not add across GPUs.

Before advertising, the node verifies pinned artifacts and tokenizer identity, starts the runtime, and checks that it produces output. Performance measurements are observations by default. Optional performance enforcement and certification tools remain available; a reference calibration campaign is not required to publish compatible choices.

Compatibility does not mean a provider is online. A published bundle can have no live offers, and live offers do not guarantee that a request will pass routing checks. Untested pairings remain estimates; provider reports describe what happened on the machine.

Run the check yourself:

```sh
katara-node bundles
katara-node bundles -all
```

The first lists what this machine can serve. The second adds the bundles it cannot, each with the reason.

## Which Models Run On What

Memory figures are the floor a pinned profile states. Compatible choices are limited to the curated set for each memory tier. New profiles assume 8,192 context tokens and one request, with reserved memory. Pairings are estimates until providers run them. Legacy identities remain available for existing installations; the curated picker uses the expanded profiles. Artifact publication and live provider availability are separate.

| Model | Artifact Status | Apple Silicon | Linux NVIDIA | Windows NVIDIA | Tools |
| --- | --- | --- | --- | --- | --- |
| Llama 3.1 8B Instruct | Pinned · existing profile | 24 GB unified | no profile | no profile | no |
| Qwen3 30B (Windows) | Pinned · existing profile | no profile | no profile | 90 GB video memory, compute 12.0 | no |
| Qwen3 4B (Windows) | Pinned · existing profile | no profile | no profile | 90 GB video memory, compute 12.0 | no |
| Qwen3-Coder 30B-A3B Instruct | Pinned · existing profile | 32 GB unified | no profile | no profile | yes |
| Qwen3 Coder 30B (Windows) | Pinned · existing profile | no profile | no profile | 90 GB video memory, compute 12.0 | yes |
| Qwen3 Next 80B (Windows) | Pinned · existing profile | no profile | no profile | 90 GB video memory, compute 12.0 | no |
| Qwen3-0.6B (GGUF) | Pinned · estimated fit | no profile | 4 GB video memory, compute 6.1 | 4 GB video memory, compute 6.1 | no |
| Qwen3-0.6B (MLX) | Pinned · estimated fit | 8 GB unified | no profile | no profile | no |
| Qwen3-4B (GGUF) | Pinned · estimated fit | no profile | 6 GB video memory, compute 6.1 | 6 GB video memory, compute 6.1 | no |
| Qwen3-8B (GGUF) | Pinned · estimated fit | no profile | 8 GB video memory, compute 6.1 | 8 GB video memory, compute 6.1 | no |
| Qwen3-8B (MLX) | Pinned · estimated fit | 16 GB unified | no profile | no profile | no |
| Qwen3-14B (GGUF) | Pinned · estimated fit | no profile | 12 GB video memory, compute 6.1 | 12 GB video memory, compute 6.1 | no |
| Qwen3-32B (GGUF) | Pinned · estimated fit | no profile | 32 GB video memory, compute 6.1 | 32 GB video memory, compute 6.1 | no |
| Qwen3-32B (MLX) | Pinned · estimated fit | 32 GB unified | no profile | no profile | no |
| Qwen3-30B-A3B (GGUF) | Pinned · estimated fit | no profile | 24 GB video memory, compute 6.1 | 24 GB video memory, compute 6.1 | no |
| Qwen3-30B-A3B (MLX) | Pinned · estimated fit | 32 GB unified | no profile | no profile | no |
| Qwen3-30B-A3B-Instruct-2507 (GGUF) | Pinned · estimated fit | no profile | 24 GB video memory, compute 6.1 | 24 GB video memory, compute 6.1 | no |
| Qwen3-Coder-30B-A3B-Instruct (GGUF) | Pinned · estimated fit | no profile | 24 GB video memory, compute 6.1 | 24 GB video memory, compute 6.1 | no |
| Qwen3-Coder-30B-A3B-Instruct (MLX) | Pinned · estimated fit | 32 GB unified | no profile | no profile | no |
| Qwen2.5-0.5B-Instruct (GGUF) | Pinned · estimated fit | no profile | 3 GB video memory, compute 6.1 | 3 GB video memory, compute 6.1 | no |
| Qwen2.5-14B-Instruct (GGUF) | Pinned · estimated fit | no profile | 16 GB video memory, compute 6.1 | 16 GB video memory, compute 6.1 | no |
| Qwen2.5-32B-Instruct (GGUF) | Pinned · estimated fit | no profile | 32 GB video memory, compute 6.1 | 32 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-0.5B-Instruct (GGUF) | Pinned · estimated fit | no profile | 3 GB video memory, compute 6.1 | 3 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-0.5B-Instruct (MLX) | Pinned · estimated fit | 8 GB unified | no profile | no profile | no |
| Qwen2.5-Coder-1.5B-Instruct (GGUF) | Pinned · estimated fit | no profile | 4 GB video memory, compute 6.1 | 4 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-7B-Instruct (GGUF) | Pinned · estimated fit | no profile | 8 GB video memory, compute 6.1 | 8 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-14B-Instruct (GGUF) | Pinned · estimated fit | no profile | 16 GB video memory, compute 6.1 | 16 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-14B-Instruct (MLX) | Pinned · estimated fit | 16 GB unified | no profile | no profile | no |
| Qwen2.5-Coder-32B-Instruct (GGUF) | Pinned · estimated fit | no profile | 32 GB video memory, compute 6.1 | 32 GB video memory, compute 6.1 | no |
| Qwen2.5-Coder-32B-Instruct (MLX) | Pinned · estimated fit | 32 GB unified | no profile | no profile | no |
| Llama-3.3-70B-Instruct (GGUF) | Pinned · estimated fit | no profile | 64 GB video memory, compute 6.1 | 64 GB video memory, compute 6.1 | no |
| Llama-3.3-70B-Instruct (MLX) | Pinned · estimated fit | 64 GB unified | no profile | no profile | no |

## Apple Silicon

The node reads the chip and memory size from the system. A model needs a matching Apple Silicon profile; memory alone does not add a chip to that profile.

| Chip | Unified Memory | GPU Cores | Bundles It Can Serve |
| --- | --- | --- | --- |
| M1 | 8 GB | 7 to 8 | Qwen2.5-Coder-0.5B-Instruct, Qwen3-0.6B |
| M1 | 16 GB | 7 to 8 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M1 Pro | 16 GB | 14 to 16 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M1 Pro | 32 GB | 14 to 16 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M1 Max | 32 GB | 24 to 32 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M1 Max | 64 GB | 24 to 32 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M1 Ultra | 64 GB, 128 GB | 48 to 64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M2 | 8 GB | 8 to 10 | Qwen2.5-Coder-0.5B-Instruct, Qwen3-0.6B |
| M2 | 16 GB, 24 GB | 8 to 10 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M2 Pro | 16 GB | 16 to 19 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M2 Pro | 32 GB | 16 to 19 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M2 Max | 32 GB | 30 to 38 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M2 Max | 64 GB, 96 GB | 30 to 38 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M2 Ultra | 64 GB, 128 GB, 192 GB | 60 to 76 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M3 | 8 GB | 8 to 10 | Qwen2.5-Coder-0.5B-Instruct, Qwen3-0.6B |
| M3 | 16 GB, 24 GB | 8 to 10 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M3 Pro | 18 GB | 14 to 18 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M3 Pro | 36 GB | 14 to 18 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M3 Max | 36 GB, 48 GB | 30 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M3 Max | 64 GB, 96 GB, 128 GB | 30 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M3 Ultra | 96 GB, 256 GB, 512 GB | 60 to 80 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M4 | 16 GB, 24 GB | 8 to 10 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M4 | 32 GB | 8 to 10 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M4 Pro | 24 GB | 16 to 20 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M4 Pro | 48 GB | 16 to 20 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M4 Pro | 64 GB | 16 to 20 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M4 Max | 36 GB, 48 GB | 32 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M4 Max | 64 GB, 128 GB | 32 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M5 | 16 GB, 24 GB | 8 to 10 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M5 | 32 GB | 8 to 10 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M5 Pro | 24 GB | 16 to 20 | Qwen2.5-Coder-14B-Instruct, Qwen3-8B |
| M5 Pro | 48 GB | 16 to 20 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M5 Pro | 64 GB | 16 to 20 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M5 Max | 36 GB, 48 GB | 32 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen3-30B-A3B |
| M5 Max | 64 GB, 128 GB | 32 to 40 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| M5 Ultra | 96 GB, 256 GB, 512 GB | 60 to 80 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |

## NVIDIA

CUDA support is specific to the pinned runtime and profile. Linux container profiles and native Windows profiles are checked separately. A card with enough video memory can still lack the required CUDA instructions.

| GPU | Memory | Compute Capability | Operating System | Compatible Models |
| --- | --- | --- | --- | --- |
| GeForce RTX 3060 | 8 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3060 | 8 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3060 | 12 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 3060 | 12 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 3080 | 10 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3080 | 10 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3080 | 12 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 3080 | 12 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 3090 | 24 GB | 8.6 | Linux amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| GeForce RTX 3090 | 24 GB | 8.6 | Windows amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| GeForce RTX 4060 Ti | 8 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4060 Ti | 8 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4060 Ti | 16 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 4060 Ti | 16 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 4070 | 12 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 4070 | 12 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 4080 | 16 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 4080 | 16 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 4090 | 24 GB | 8.9 | Linux amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| GeForce RTX 4090 | 24 GB | 8.9 | Windows amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| GeForce RTX 5070 | 12 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 5070 | 12 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 5080 | 16 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 5080 | 16 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 5090 | 32 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| GeForce RTX 5090 | 32 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX A4000 | 16 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| RTX A4000 | 16 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| RTX A5000 | 24 GB | 8.6 | Linux amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| RTX A5000 | 24 GB | 8.6 | Windows amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| RTX A6000 | 48 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX A6000 | 48 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX 4000 Ada | 20 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| RTX 4000 Ada | 20 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| RTX 5000 Ada | 32 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX 5000 Ada | 32 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX 6000 Ada | 48 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX 6000 Ada | 48 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| RTX PRO 6000 Blackwell | 96 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| RTX PRO 6000 Blackwell | 96 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| A30 | 24 GB | 8.0 | Linux amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| A30 | 24 GB | 8.0 | Windows amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| A100 | 40 GB | 8.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| A100 | 40 GB | 8.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| A100 | 80 GB | 8.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| A100 | 80 GB | 8.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| L4 | 24 GB | 8.9 | Linux amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| L4 | 24 GB | 8.9 | Windows amd64 | Qwen3-Coder-30B-A3B-Instruct, Qwen3-30B-A3B, Qwen3-30B-A3B-Instruct-2507 |
| L40S | 48 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| L40S | 48 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Qwen3-32B, Qwen2.5-32B-Instruct |
| H100 | 80 GB | 9.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| H100 | 80 GB | 9.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| H200 | 141 GB | 9.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| H200 | 141 GB | 9.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| B200 | 180 GB | 10.0 | Linux amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| B200 | 180 GB | 10.0 | Windows amd64 | Qwen2.5-Coder-32B-Instruct, Qwen3-Coder-30B-A3B-Instruct, Llama-3.3-70B-Instruct |
| GeForce RTX 4060 Laptop GPU | 8 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4060 Laptop GPU | 8 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4060 | 8 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4060 | 8 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3050 | 6 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 3050 | 6 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 3050 | 8 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3050 | 8 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 | 8 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 | 8 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 Laptop GPU | 8 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 Laptop GPU | 8 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 Ti | 8 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 Ti | 8 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 5060 Ti | 16 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 5060 Ti | 16 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce GTX 1650 | 4 GB | 7.5 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-0.6B |
| GeForce GTX 1650 | 4 GB | 7.5 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-0.6B |
| GeForce RTX 3060 Ti | 8 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3060 Ti | 8 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3070 | 8 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3070 | 8 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3060 Laptop GPU | 6 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 3060 Laptop GPU | 6 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 5070 Ti | 16 GB | 12.0 | Linux amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 5070 Ti | 16 GB | 12.0 | Windows amd64 | Qwen2.5-Coder-14B-Instruct, Qwen3-14B, Qwen2.5-14B-Instruct |
| GeForce RTX 4050 Laptop GPU | 6 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 4050 Laptop GPU | 6 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 2060 | 6 GB | 7.5 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 2060 | 6 GB | 7.5 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 2060 | 12 GB | 7.5 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 2060 | 12 GB | 7.5 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce GTX 1060 | 3 GB | 6.1 | Linux amd64 | Qwen2.5-Coder-0.5B-Instruct, Qwen2.5-0.5B-Instruct |
| GeForce GTX 1060 | 3 GB | 6.1 | Windows amd64 | Qwen2.5-Coder-0.5B-Instruct, Qwen2.5-0.5B-Instruct |
| GeForce GTX 1060 | 6 GB | 6.1 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce GTX 1060 | 6 GB | 6.1 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce RTX 4070 SUPER | 12 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 4070 SUPER | 12 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce GTX 1660 SUPER | 6 GB | 7.5 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce GTX 1660 SUPER | 6 GB | 7.5 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce GTX 1050 Ti | 4 GB | 6.1 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-0.6B |
| GeForce GTX 1050 Ti | 4 GB | 6.1 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-0.6B |
| GeForce RTX 4070 Laptop GPU | 8 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4070 Laptop GPU | 8 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 4070 Ti | 12 GB | 8.9 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 4070 Ti | 12 GB | 8.9 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-14B |
| GeForce RTX 3070 Ti | 8 GB | 8.6 | Linux amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce RTX 3070 Ti | 8 GB | 8.6 | Windows amd64 | Qwen2.5-Coder-7B-Instruct, Qwen3-8B |
| GeForce GTX 1660 Ti | 6 GB | 7.5 | Linux amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |
| GeForce GTX 1660 Ti | 6 GB | 7.5 | Windows amd64 | Qwen2.5-Coder-1.5B-Instruct, Qwen3-4B |

### Cards Below The Floor

Intel Iris Xe Graphics: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

AMD Radeon Graphics: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

AMD Radeon RX 9070 XT: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

Intel UHD Graphics: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

AMD Radeon(TM) Graphics: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

AMD Radeon RX 6600: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

AMD Radeon RX 9060 XT: The provider has no supported AMD or Intel GPU runtime. Shared system memory is not dedicated NVIDIA VRAM.

## Reading The Memory Numbers

Two things trip people up.

A card is sold by its nominal size and reports something slightly smaller. A 24 GB board reports about 24564 MiB once the driver has taken its share, which is under 24 GiB. Bundle floors sit below the nominal size for that reason, so a card sold as 24 GB clears a floor written for 24 GB cards. Apple reports the exact physical size, so unified memory floors are compared as written.

Memory does not add up across cards. One model runs in one process on one card, so a profile has to fit the largest single card in the machine. Two 24 GB cards will not serve a bundle that asks for 46 GB.

## Operating Systems

Apple Silicon profiles use macOS arm64 and the pinned MLX runtime. Linux NVIDIA profiles use their pinned container, Docker and the NVIDIA container toolkit. Native Windows amd64 NVIDIA profiles use pinned llama.cpp executables and CUDA dependencies. An available profile for one operating system does not establish support on another.

AMD and Intel GPU backends are not currently supported by the provider. Shared system memory must not be entered as dedicated NVIDIA video memory.

## Building Your Own Check

Everything in the tables above is published as JSON at `https://docs.katara.com/hardware.json`: the device tables, and every bundle's hardware floors read from the published manifests. It is generated with the site, so a new bundle or a newly listed card appears without anyone editing a page. The endpoint allows cross-origin reads.

The matching rules are published next to it at `https://docs.katara.com/hardware-check.js`, an ES module with the browser detection and the same comparison the node makes. Note what a browser will not tell you: no browser reports video memory, and none reports the size of a Mac's unified memory. Chromium names the GPU, Safari reports a flat `Apple GPU` and nothing more. Any honest check in a browser detects the chip and asks for the memory.

The node checks the real machine. `katara-node bundles` uses the same catalog policy and verifies the selected manifest before listing a choice. Artifact hashes and signed identities remain required. A provider report is an observation, not proof that every request executed the advertised weights.