Can I run Qwen3.8 27B on the GeForce RTX 5060 Ti 16GB?
Runs slowlyQ4_K_M does not fit fully in VRAM, so 3.2 GB of the weights run from system RAM: Runs slowly, est. 8.4 tok/s (6.7–10.0, calibrated estimate ±20%).
16 GB VRAM · 448.0 GB/s · FP16 23.7 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache. The verdict and quant file above stay the same anywhere from 16 to 128 GB of system RAM.
Every quant of Qwen3.8 27B on the GeForce RTX 5060 Ti 16GB
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| Q2_K | 10.82 GB | Runs well | est. 27.6 tok/s24.3–30.9calibrated estimate ±12% | 11.7 / 16.0 GB | Full GPU | — |
| IQ4_XS | 15.48 GB | Runs well | est. 13.4 tok/s10.7–16.1calibrated estimate ±20% | 16.0 / 16.0 GB + 1.2 GB RAM | Partial offload | — |
| Q4_K_Mbaseline | 17.44 GB | Runs slowly | est. 8.4 tok/s6.7–10.0calibrated estimate ±20% | 16.0 / 16.0 GB + 3.2 GB RAM | Partial offload | — |
| Q8_0 | 29.12 GB | Won't run | est. 2.6 tok/s2.1–3.1calibrated estimate ±20% | 16.0 / 16.0 GB + 14.8 GB RAM | Partial offload | Fits, but too slow to use |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Where the memory goes at Q4_K_M
- Weights
- 14.3 GB
- KV cache
- 0.5 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Weights in system RAM
- 3.2 GB
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 17.3 GBRuns slowly est. 8.9 tok/s7.2–10.7calibrated estimate ±20% | 17.2 GBRuns slowly est. 9.2 tok/s7.4–11.1calibrated estimate ±20% | 17.1 GBRuns slowly est. 9.4 tok/s7.5–11.2calibrated estimate ±20% |
| 8k | 17.6 GBRuns slowly est. 8.4 tok/s6.7–10.0calibrated estimate ±20% | 17.3 GBRuns slowly est. 8.9 tok/s7.1–10.6calibrated estimate ±20% | 17.2 GBRuns slowly est. 9.1 tok/s7.3–11.0calibrated estimate ±20% |
| 16k | 18.2 GBRuns slowly est. 7.4 tok/s5.9–8.9calibrated estimate ±20% | 17.7 GBRuns slowly est. 8.2 tok/s6.6–9.8calibrated estimate ±20% | 17.4 GBRuns slowly est. 8.7 tok/s7.0–10.4calibrated estimate ±20% |
| 32k | 19.4 GBRuns slowly est. 5.9 tok/s4.7–7.1calibrated estimate ±20% | 18.4 GBRuns slowly est. 7.1 tok/s5.7–8.5calibrated estimate ±20% | 17.9 GBRuns slowly est. 7.9 tok/s6.4–9.5calibrated estimate ±20% |
| 64k | 21.9 GBRuns slowly est. 4.1 tok/s3.2–4.9calibrated estimate ±20% | 19.9 GBRuns slowly est. 5.5 tok/s4.4–6.6calibrated estimate ±20% | 18.8 GBRuns slowly est. 6.7 tok/s5.4–8.1calibrated estimate ±20% |
| 128k | 26.8 GBWon't run est. 2.3 tok/s1.8–2.8calibrated estimate ±20% | 22.8 GBRuns slowly est. 3.7 tok/s2.9–4.4calibrated estimate ±20% | 20.6 GBRuns slowly est. 5.0 tok/s4.0–6.0calibrated estimate ±20% |
| 256k | 36.5 GBdoes not fit | 28.6 GBWon't run est. 2.0 tok/s1.6–2.4calibrated estimate ±20% | 24.3 GBRuns slowly est. 3.2 tok/s2.6–3.8calibrated estimate ±20% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs slowly up to 64k context with an f16 KV cache, and up to 128k with q8_0.
Estimated speed
- Generation
- est. 8.4 tok/s (6.7–10.0, calibrated estimate ±20%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Measured on this exact combination
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Hardware | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| GeForce RTX 5060 Ti 16GB | UD-IQ4_XS | llama.cpp | 8k | — | 27.6 | KV q8_0 | github.com | 2026-10-11 |
If this is not enough
A smaller model that runs well on the GeForce RTX 5060 Ti 16GB
Runs wellGemma 4 26B-A4B: Runs well, est. 59.2 tok/s (41.4–76.9, theoretical estimate ±30%)
A lower quant of Qwen3.8 27B
Runs wellIQ4_XS (15.48 GB): Runs well, est. 13.4 tok/s (10.7–16.1, calibrated estimate ±20%)
Cheapest GPU that runs Qwen3.8 27B great
Runs greatGeForce RTX 3090 — $1,050 street (as of 2026-08-09): est. 36.4 tok/s (32.1–40.8, calibrated estimate ±12%).
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen3.8-27B-Q4_K_M.gguf (17.44 GB) from bartowski/Qwen3.8-27B-GGUF.
llama-server -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M -c 8192 -ngl 52 -fa on- -ngl 52 puts 52 of 64 layers on the GPU; if VRAM overflows, lower it by 1–2.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 0.29 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
- Ollama splits layers between GPU and CPU automatically.
/set parameter num_ctx 8192
Related pages
Guide: How much VRAM do local LLMs need? GPU requirements for 8–32 GB (2026)
Other models on the GeForce RTX 5060 Ti 16GB
Qwen3.8 27B on other hardware
Frequently asked questions
Can I run Qwen3.8 27B on the GeForce RTX 5060 Ti 16GB?
Q4_K_M does not fit fully in VRAM, so 3.2 GB of the weights run from system RAM: Runs slowly, est. 8.4 tok/s (6.7–10.0, calibrated estimate ±20%). Q4_K_M needs 17.6 GB at 8k context; this setup has 15.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can Qwen3.8 27B use on the GeForce RTX 5060 Ti 16GB?
At Q4_K_M the verdict stays Runs slowly up to 64k context with an f16 KV cache (4.3 GB of KV) and up to 128k with a q8_0 KV cache (4.6 GB). The model supports up to 256k.
Which quant should I use, and how fast is it?
Q4_K_M (17.44 GB) is the recommended balance: Runs slowly, est. 8.4 tok/s (6.7–10.0, calibrated estimate ±20%). It is also the largest tracked file that gets this verdict.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.