Best local LLMs for the RX 9060 XT (17 GB) — measured, not guessed

Q: Do I need to install CUDA, ROCm or Python to use the RX 9060 XT?

No. npx turbollm detects your GPU and auto-provisions a matching prebuilt llama-server — no toolkit, no Python environment, no compiler. Community forks are available too, from the engines catalog.

Most "best models for your GPU" pages are guesswork. This one is a readout: TurboLLM users running this exact card have contributed 73 real benchmarks across 3 machines, covering 9 model/quant configurations. Users have pushed it to a 256K-token context on Qwen3.8 27B Heretic Ara. Every number below carries its own sample size.

Numbers you can trust — including the small ones

Nothing here is extrapolated from another GPU or read off a spec sheet. These are medians of opt-in benchmark telemetry from real TurboLLM installs on 17 GB of VRAM, aggregated over 90 days. Some configurations have been measured by a single machine — we show you that instead of hiding it. And when you run TurboLLM yourself, it benchmarks the model on your card and shows a VRAM-fit verdict before you download.

Measured on the RX 9060 XT

Model · quant	Median tok/s	Context	Measured during	Sample
Qwen3.8 27B Heretic Ara _{IQ3_M}	16 t/s	256K	real use	_{n=23 · 1 machine}
Gemma 4 26B A4B It Qat Q4_0 Unquantized Heretic _{Q4_0 · MoE}	31 t/s	256K	real use	_{n=14 · 1 machine}
Ornith Aeon Bf16 _{Q4_K_M · MoE}	10 t/s	256K	real use	_{n=6 · 1 machine}
Qwen3.6 35B A3B Uncensored Heretic _{Q4_K_M · MoE}	24 t/s	256K	real use	_{n=4 · 1 machine}
Ternary Bonsai 27B Unpacked Heretic Ja _{Q2_0}	32 t/s	256K	real use	_{n=3 · 1 machine}
Huihui Gemma 3n E4B It Abliterated _{Q4_K_M}	65 t/s	32K	real use	_{n=2 · 1 machine}
Gemma 4 12B Agentic Fable5 v2 _{Q4_K_M}	25 t/s	256K	real use	_{n=2 · 1 machine}
Lordx64 Distill _{BF16 · MoE}	22 t/s	256K	real use	_{n=2 · 1 machine}
Ornith 1.5 9B _{Q8_0}	9 t/s	256K	real use	_{n=2 · 1 machine}

_{Medians of real, opt-in benchmark telemetry from TurboLLM installs on this GPU — not a
spec sheet, not an estimate. Every row shows its own sample size; small
samples are labeled rather than hidden. Auto-tune sweep rows are measured on an
idle machine with a cold cache; real use rows are measured during actual chat,
gateway and coding traffic. The two are never averaged together. Context is the largest
window that configuration was actually run at, not a theoretical maximum.}

How TurboLLM fits a model to this card

Load any model and TurboLLM benchmarks it on your own RX 9060 XT, storing the measured tokens/sec in the model list — the same measurements that, with your consent, become rows in the table above. Before you download, every quant gets a VRAM-fit verdict against your real free VRAM, including KV-cache growth at your chosen context. Flip the auto-fit toggle and it picks the GPU-layer count and, for Mixture-of-Experts models, the expert-offload split — and it decides fit by measuring host-memory spill directly, because used-VRAM alone silently reads as a clean fit while gigabytes run from system RAM. Details in the auto-tune docs.

Want your card on this page?

This page exists because TurboLLM users opted into anonymous benchmark sharing. Turn it on in Settings and your measurements join the dataset — no account, no identity, no prompt content, just the model, the quant, the hardware and the speed. The more people share, the better every one of these pages gets.

FAQ

What LLMs can you actually run on the RX 9060 XT?

TurboLLM users on the RX 9060 XT have benchmarked 9 model/quant configurations across 3 machines, on 17 GB of VRAM. The fastest measured so far is Huihui Gemma 3n E4B It Abliterated at Q4_K_M, with a median of 65 tokens/sec. Every figure on this page comes from real runs on this card, with its sample size shown — nothing is extrapolated from a different GPU.

How fast is Huihui Gemma 3n E4B It Abliterated on the RX 9060 XT?

Median 65 tokens/sec at Q4_K_M, from 2 runs across 1 machine running TurboLLM. Your own figure depends on context length, quant and settings, which is why TurboLLM re-measures on your machine every time you load a model instead of quoting someone else's number.

Where do these numbers come from?

Opt-in, anonymous benchmark telemetry from TurboLLM installs — the same measurements the app shows you locally after a load. They are aggregated over a 90-day window and published with per-row sample sizes. Auto-tune sweeps (idle machine, cold cache) are kept separate from measurements taken during real chat, gateway and coding traffic, so a best-case number is never mixed into a real-world one.

Do I need to install CUDA, ROCm or Python to use the RX 9060 XT?

No. npx turbollm detects your GPU and auto-provisions a matching prebuilt llama-server — no toolkit, no Python environment, no compiler. Community forks are available too, from the engines catalog.

Run it on your RX 9060 XT

$ npx turbollm

One command: it detects your GPU, provisions a matching engine, and opens the UI. Already have GGUF files in an LM Studio folder? TurboLLM indexes them in place — no re-download. New here? Start with Install & first run and Quantization explained, or browse the full model hub.