Can I run Qwen3.6 35B-A3B on the GeForce RTX 5060 Ti 16GB?
Runs wellQ4_K_M keeps attention and the experts of 25 layers on the GPU (15.1 GB) and the remaining experts in 8.5 GB of system RAM: Runs well, est. 69.5 tok/s (48.7–90.4, theoretical estimate ±30%). Q2_K (heavy quality loss) would reach Runs great, est. 120.3 tok/s (96.2–144.3, 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 Qwen3.6 35B-A3B on the GeForce RTX 5060 Ti 16GB
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| Q2_K | 13.51 GB | Runs great | est. 120.3 tok/s96.2–144.3calibrated estimate ±20% | 13.8 / 16.0 GB | Full GPU | — |
| IQ4_XS | 19.70 GB | Runs well | est. 83.2 tok/s58.3–108.2theoretical estimate ±30% | 15.2 / 16.0 GB + 5.9 GB RAM | MoE experts in RAM | Not fully on the GPU — Runs great needs the whole model in GPU memory |
| Q4_K_Mbaseline | 22.29 GB | Runs well | est. 69.5 tok/s48.7–90.4theoretical estimate ±30% | 15.1 / 16.0 GB + 8.5 GB RAM | MoE experts in RAM | Not fully on the GPU — Runs great needs the whole model in GPU memory |
| Q8_0 | 37.81 GB | Runs well | est. 37.0 tok/s25.9–48.1theoretical estimate ±30% | 14.6 / 16.0 GB + 24.5 GB RAM | MoE experts in RAM | Not fully on the GPU — Runs great needs the whole model in GPU memory |
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
- 13.8 GB
- KV cache
- 0.2 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 0.9 GB
- Weights in system RAM
- 8.5 GB
The system RAM share above also counts the 0.90 GB MTP (multi-token prediction) block, which llama.cpp does not load by default, so it errs on the safe side. Turning on MTP speculative decoding (llama.cpp --spec-type draft-mtp) loads that block and needs KV cache for it too.
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 21.7 GBRuns well est. 72.3 tok/s50.6–94.0theoretical estimate ±30% | 21.7 GBRuns well est. 73.0 tok/s51.1–94.9theoretical estimate ±30% | 21.7 GBRuns well est. 73.3 tok/s51.3–95.3theoretical estimate ±30% |
| 8k | 21.8 GBRuns well est. 69.5 tok/s48.7–90.4theoretical estimate ±30% | 21.8 GBRuns well est. 70.7 tok/s49.5–92.0theoretical estimate ±30% | 21.7 GBRuns well est. 72.9 tok/s51.1–94.8theoretical estimate ±30% |
| 16k | 22.1 GBRuns well est. 67.0 tok/s46.9–87.1theoretical estimate ±30% | 21.9 GBRuns well est. 69.3 tok/s48.5–90.1theoretical estimate ±30% | 21.8 GBRuns well est. 70.6 tok/s49.5–91.8theoretical estimate ±30% |
| 32k | 22.6 GBRuns well est. 61.4 tok/s43.0–79.8theoretical estimate ±30% | 22.2 GBRuns well est. 65.4 tok/s45.8–85.0theoretical estimate ±30% | 22.1 GBRuns well est. 69.2 tok/s48.4–89.9theoretical estimate ±30% |
| 64k | 23.5 GBRuns well est. 52.6 tok/s36.8–68.4theoretical estimate ±30% | 22.9 GBRuns well est. 59.8 tok/s41.9–77.7theoretical estimate ±30% | 22.6 GBRuns well est. 65.1 tok/s45.6–84.6theoretical estimate ±30% |
| 128k | 25.5 GBRuns well est. 40.9 tok/s28.6–53.2theoretical estimate ±30% | 24.2 GBRuns well est. 50.3 tok/s35.2–65.4theoretical estimate ±30% | 23.6 GBRuns well est. 58.3 tok/s40.8–75.7theoretical estimate ±30% |
| 256k | 29.4 GBRuns slowly est. 28.3 tok/s19.8–36.8theoretical estimate ±30% | 26.9 GBRuns well est. 38.6 tok/s27.0–50.1theoretical estimate ±30% | 25.5 GBRuns well est. 48.1 tok/s33.7–62.6theoretical estimate ±30% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs well up to 128k context with an f16 KV cache, and up to 256k with q8_0.
Estimated speed
- Generation
- est. 69.5 tok/s (48.7–90.4, theoretical estimate ±30%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Why this verdict: Not fully on the GPU — Runs great needs the whole model in GPU memory
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A larger model that also runs well on the GeForce RTX 5060 Ti 16GB
Runs wellK2-Horizon MoVA 36B-A4B: Runs well, est. 30.7 tok/s (21.5–39.9, theoretical estimate ±30%)
A lower quant of Qwen3.6 35B-A3B
Runs greatQ2_K (13.51 GB, heavy quality loss): Runs great, est. 120.3 tok/s (96.2–144.3, calibrated estimate ±20%)
Keep only the experts in system RAM
--n-cpu-moe 15 keeps only the experts of the first 15 layers in system RAM and everything else on the GPU (this estimate).
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen_Qwen3.6-35B-A3B-Q4_K_M.gguf (22.29 GB) from bartowski/Qwen_Qwen3.6-35B-A3B-GGUF.
llama-server -hf bartowski/Qwen_Qwen3.6-35B-A3B-GGUF:Q4_K_M -c 8192 -ngl all --n-cpu-moe 15 -fa on- -ngl all loads every layer on the GPU.
- --n-cpu-moe 15 keeps the experts in the first 15 layers in system RAM; the remaining 25 layers with experts keep them on the GPU. This estimate assumes the same split.
- If loading fails with an out-of-memory error, raise the value to 16, or to 17 if that isn't enough.
- llama.cpp b7410 (December 2025) and later split the model in a similar way on their own if you leave out -ngl and --n-cpu-moe.
- -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 0.09 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/bartowski/Qwen_Qwen3.6-35B-A3B-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 0.30 and later automatically keep part of the experts in system RAM to fit your VRAM. Older versions split whole layers instead, which is much slower.
/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
Qwen3.6 35B-A3B on other hardware
Frequently asked questions
Can I run Qwen3.6 35B-A3B on the GeForce RTX 5060 Ti 16GB?
Q4_K_M keeps attention and the experts of 25 layers on the GPU (15.1 GB) and the remaining experts in 8.5 GB of system RAM: Runs well, est. 69.5 tok/s (48.7–90.4, theoretical estimate ±30%). Q2_K (heavy quality loss) would reach Runs great, est. 120.3 tok/s (96.2–144.3, calibrated estimate ±20%). Q4_K_M needs 21.8 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 Qwen3.6 35B-A3B use on the GeForce RTX 5060 Ti 16GB?
At Q4_K_M the verdict stays Runs well up to 128k context with an f16 KV cache (2.7 GB of KV) and up to 256k with a q8_0 KV cache (2.9 GB). The model supports up to 256k.
Which quant should I use, and how fast is it?
Q4_K_M (22.29 GB) is the recommended balance: Runs well, est. 69.5 tok/s (48.7–90.4, theoretical estimate ±30%). The largest tracked file that still gets Runs well is Q8_0 (37.81 GB), est. 37.0 tok/s (25.9–48.1, theoretical estimate ±30%).
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.