Radeon RX 6800 XT — what LLMs can it run?
59 of 110 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Radeon RX 6800 XT
16 GB, 512 GB/s GDDR6. 59 of 110 models in this index fit comfortably at 4K context; 67 load at all.
- Biggest that fits
- ERNIE 4.5 21B-A3B — 14.1 GB at Q4_K_M — 1.9 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- Stable LM 2 12B Chat — 8.3 GB at Q4_K_M — under 60% of the card, which leaves 7.7 GB for a long context window or a second process.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 512 GB/s puts a hard ceiling near 111 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
- The wall you will hit
- Backend coverage, not capacity. At 512 GB/s the memory system is competitive with NVIDIA cards of the same size; what varies is whether your runtime has a working ROCm build for your kernel and card. llama.cpp and vLLM both ship official ROCm support — this is a setup question, not a hardware ceiling.
32B · 2
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| ERNIE 4.5 21B-A3B21B-A3B | Q4_K_M | 14.05 GB | +1.9 GB | ≤296Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +3.2 GB | ≤259Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q4_K_M | 13.43 GB | +2.6 GB | 26–35Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q4_K_M | 10.08 GB | +5.9 GB | ≤371Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +5.9 GB | ≤371Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +5.8 GB | 35–46Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +4.4 GB | 30–40Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +4.3 GB | 30–40Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +5.9 GB | 36–47Derived |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +4.2 GB | 30–40Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +3.1 GB | 28–37Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +4.9 GB | 32–42Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +6.2 GB | 37–49Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +2.4 GB | ≤203Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +7.7 GB | 43–57Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +3.1 GB | 27–36Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +8.4 GB | 49–65Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +8.7 GB | 50–67Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +4.3 GB | 32–43Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +5.6 GB | 32–43Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +5.8 GB | 33–44Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +5.6 GB | 34–45Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +5.6 GB | 34–45Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +6 GB | 35–47Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +10.5 GB | 63–84Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +10.1 GB | 64–85Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +8.4 GB | 47–63Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +6.5 GB | 37–49Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +10.4 GB | 65–86Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +10.4 GB | 65–86Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +10.4 GB | 65–86Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +9.5 GB | 56–74Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +10.6 GB | 68–90Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +9.2 GB | 51–67Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +10.9 GB | 69–91Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +10.9 GB | 69–91Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +7.5 GB | 40–53Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +5.7 GB | 41–54Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +5.7 GB | 41–54Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +10.9 GB | 72–96Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +9.3 GB | 54–71Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +9.3 GB | 54–71Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +10.8 GB | ≤109Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +11 GB | ≤117Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +11 GB | ≤118Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +12 GB | ≤164Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +11.2 GB | ≤132Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +10.1 GB | ≤133Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +11.6 GB | ≤139Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +12 GB | ≤158Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +12.4 GB | ≤164Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +12.3 GB | ≤165Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +12.7 GB | ≤195Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +13.8 GB | ≤259Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +13.6 GB | ≤294Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +14.2 GB | ≤330Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +14.5 GB | ≤410Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +15.1 GB | ≤632Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +15.4 GB | ≤1045Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
A further 8 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.
Measured on this card
No benchmark runs in this index were recorded on a Radeon RX 6800 XT. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.
Cards with the same budget
What fits is decided by memory, so every 16GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.
The same 16GB in system memory is a different machine — everything on this page fits there, at a speed set by your DIMMs rather than a graphics bus: 16 GB RAM (CPU)
- Radeon RX 9070 XT
- Radeon RX 9070
- Radeon RX 9060 XT 16G
- Radeon RX 7900 GRE
- Radeon RX 7800 XT
- Radeon RX 7600 XT
- Radeon RX 6900 XT
Same list, different speed — play one sentence on both cards: Radeon RX 6800 XT vs Radeon RX 9070 XT (Qwen3 14B Instruct) · Radeon RX 6800 XT vs Radeon RX 9070 (Qwen3 14B Instruct) · Radeon RX 6800 XT vs Radeon RX 9060 XT 16G (Qwen3 14B Instruct)
Setup guides for this card
Step-by-step guides written for the Radeon RX 6800 XT's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A Radeon RX 7900 XT (20GB) fits 14 more of the indexed models than this card. Radeon RX 7900 XT →
Before buying, hear the speed difference on a model both cards hold: Radeon RX 6800 XT vs Radeon RX 7900 XT (Qwen3 14B Instruct) →
Common questions
What is the best local LLM for a Radeon RX 6800 XT?
For everyday use, Stable LM 2 12B Chat at Q4_K_M — about 8.3 GB of the card's 16 GB at 4K context, leaving 7.7 GB for a longer window. If you want the largest thing that will load, that is ERNIE 4.5 21B-A3B at Q4_K_M (14.1 GB, 1.9 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can a Radeon RX 6800 XT run Codestral 22B?
Only just. Its smallest build here, Q4_K_M, needs about 15.0 GB against 16.0 GB usable — that loads with nothing else running, with no margin for a longer context window. It is not on the list above, which requires a model to stay inside 88% of usable capacity.
How many tokens per second does a Radeon RX 6800 XT do on an 8B model at Q4?
This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 512 GB/s — a ceiling near 111 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
Radeon RX 6800 XT or RTX 5080 for local LLMs?
They do not hold the same set: Radeon RX 6800 XT fits 59 and RTX 5080 63 of 110 at 4K — the cards load different formats. Speed is the other difference — 512 GB/s against 960 GB/s, a 1.9× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5080 generates faster on any model both can hold.
59 of 110 indexed models fit comfortably in 16GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml