Can I run Qwen3.6 35B-A3B on the GeForce RTX 5070?
Runs wellQ4_K_M keeps attention and the experts of 17 layers on the GPU (11.2 GB) and the remaining experts in 12.4 GB of system RAM: Runs well, est. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%).
12 GB VRAM · 672.0 GB/s · FP16 30.9 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache. The GeForce RTX 5070 Ti Laptop has the same memory setup (12 GB, 672.0 GB/s), so its generation-speed estimate is the same; only prompt processing differs.
Verdict by system RAM
This page assumes 32 GB of system RAM. For this pair, the verdict depends on how much system RAM you have.
- 16 GBWon't runTry IQ4_XS: Runs wellest. 80.9 tok/s (56.7–105.2, theoretical estimate ±30%)
- 32 GBRuns wellest. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%)(used on this page)
- 64 GBRuns wellest. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%)
- 96 GBRuns wellest. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%)
- 128 GBRuns wellest. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%)
Every quant of Qwen3.6 35B-A3B on the GeForce RTX 5070
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| Q2_K | 13.51 GB | Runs well | est. 128.7 tok/s90.1–167.4theoretical estimate ±30% | 11.6 / 12.0 GB + 3.3 GB RAM | MoE experts in RAM | Not fully on the GPU — Runs great needs the whole model in GPU memory |
| IQ4_XS | 19.70 GB | Runs well | est. 80.9 tok/s56.7–105.2theoretical estimate ±30% | 11.3 / 12.0 GB + 9.7 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. 68.6 tok/s48.0–89.2theoretical estimate ±30% | 11.2 / 12.0 GB + 12.4 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 | Won't run | — | Needs about 37.1 GB; 11.4 GB of VRAM and 28.0 GB of RAM are free | — | — |
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
- 9.9 GB
- KV cache
- 0.2 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 0.8 GB
- Weights in system RAM
- 12.4 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.
System RAM, context length and KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 21.7 GBRuns well est. 69.5 tok/s48.6–90.3theoretical estimate ±30% | 21.7 GBRuns well est. 69.9 tok/s48.9–90.8theoretical estimate ±30% | 21.7 GBRuns well est. 70.1 tok/s49.1–91.1theoretical estimate ±30% |
| 8k | 21.8 GBRuns well est. 68.6 tok/s48.0–89.2theoretical estimate ±30% | 21.8 GBRuns well est. 69.4 tok/s48.6–90.2theoretical estimate ±30% | 21.7 GBRuns well est. 69.8 tok/s48.9–90.8theoretical estimate ±30% |
| 16k | 22.1 GBRuns well est. 67.0 tok/s46.9–87.1theoretical estimate ±30% | 21.9 GBRuns well est. 68.5 tok/s48.0–89.1theoretical estimate ±30% | 21.8 GBRuns well est. 69.4 tok/s48.5–90.2theoretical estimate ±30% |
| 32k | 22.6 GBRuns well est. 62.7 tok/s43.9–81.5theoretical estimate ±30% | 22.2 GBRuns well est. 65.4 tok/s45.8–85.0theoretical estimate ±30% | 22.1 GBRuns well est. 68.4 tok/s47.9–88.9theoretical estimate ±30% |
| 64k | 23.5 GBRuns well est. 55.6 tok/s38.9–72.3theoretical estimate ±30% | 22.9 GBRuns well est. 61.1 tok/s42.8–79.5theoretical estimate ±30% | 22.6 GBRuns well est. 65.2 tok/s45.6–84.8theoretical estimate ±30% |
| 128k | 25.5 GBRuns well est. 45.3 tok/s31.7–58.9theoretical estimate ±30% | 24.2 GBRuns well est. 53.2 tok/s37.2–69.1theoretical estimate ±30% | 23.6 GBRuns well est. 59.6 tok/s41.8–77.5theoretical estimate ±30% |
| 256k | 29.4 GBRuns well est. 33.1 tok/s23.1–43.0theoretical estimate ±30% | 26.9 GBRuns well est. 42.7 tok/s29.9–55.6theoretical estimate ±30% | 25.5 GBRuns well est. 51.0 tok/s35.7–66.3theoretical 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 256k context with an f16 KV cache, and up to 256k with q8_0.
Estimated speed
- Generation
- est. 68.6 tok/s (48.0–89.2, 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
Keep only the experts in system RAM
--n-cpu-moe 23 keeps only the experts of the first 23 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.
These commands assume 32 GB of system RAM. For a different amount of RAM, pick it under “System RAM, context length and KV cache” above to see the commands for it.
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 23 -fa on- -ngl all loads every layer on the GPU.
- --n-cpu-moe 23 keeps the experts in the first 23 layers in system RAM; the remaining 17 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 24, or to 25 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 5070
Qwen3.6 35B-A3B on other hardware
Frequently asked questions
Can I run Qwen3.6 35B-A3B on the GeForce RTX 5070?
Q4_K_M keeps attention and the experts of 17 layers on the GPU (11.2 GB) and the remaining experts in 12.4 GB of system RAM: Runs well, est. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%). Q4_K_M needs 21.8 GB at 8k context; this setup has 11.4 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 5070?
At Q4_K_M the verdict stays Runs well up to 256k context with an f16 KV cache (5.4 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. 68.6 tok/s (48.0–89.2, theoretical estimate ±30%). 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.