Can I run Qwen3.5 9B on the GeForce RTX 5070?
Runs greatQ4_K_M fits in VRAM with 5.4 GB to spare: Runs great, est. 76.6 tok/s (67.4–85.8, calibrated estimate ±12%).
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 verdict and quant file above stay the same anywhere from 16 to 128 GB of system RAM. 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. Prompt processing: about 3,375 tok/s (2,025–4,724, rough estimate ±40%) here vs about 3,276 tok/s (1,966–4,587, rough estimate ±40%) on the GeForce RTX 5070 Ti Laptop.
Every quant of Qwen3.5 9B on the GeForce RTX 5070
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| IQ4_XS | 5.47 GB | Runs great | est. 82.0 tok/s72.1–91.8calibrated estimate ±12% | 6.2 / 12.0 GB | Full GPU | — |
| Q4_K_Mbaseline | 5.87 GB | Runs great | est. 76.6 tok/s67.4–85.8calibrated estimate ±12% | 6.6 / 12.0 GB | Full GPU | — |
| Q8_0 | 9.79 GB | Runs great | 67.3 tok/s8k estimate 46.8 tok/s (41.2–52.4, calibrated estimate ±12%)measured (1 run, 32k context) | 9.9 / 12.0 GB | Full GPU | — |
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
- 5.1 GB
- KV cache
- 0.3 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 5.4 GB
0.75 GB of the 5.87 GB file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM and does not load the 0.17 GB MTP (multi-token prediction) block by default. Turning on MTP speculative decoding (llama.cpp --spec-type draft-mtp) needs memory for that block plus KV cache for it.
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 5.8 GBRuns great est. 78.3 tok/s68.9–87.7calibrated estimate ±12% | 5.7 GBRuns great est. 79.2 tok/s69.7–88.7calibrated estimate ±12% | 5.7 GBRuns great est. 79.6 tok/s70.1–89.2calibrated estimate ±12% |
| 8k | 6.0 GBRuns great est. 76.6 tok/s67.4–85.8calibrated estimate ±12% | 5.8 GBRuns great est. 78.2 tok/s68.8–87.6calibrated estimate ±12% | 5.8 GBRuns great est. 79.1 tok/s69.6–88.6calibrated estimate ±12% |
| 16k | 6.3 GBRuns great est. 73.4 tok/s64.6–82.2calibrated estimate ±12% | 6.1 GBRuns great est. 76.4 tok/s67.2–85.6calibrated estimate ±12% | 5.9 GBRuns great est. 78.1 tok/s68.7–87.5calibrated estimate ±12% |
| 32k | 7.0 GBRuns great est. 67.7 tok/s59.6–75.9calibrated estimate ±12% | 6.5 GBRuns great est. 73.0 tok/s64.2–81.8calibrated estimate ±12% | 6.2 GBRuns great est. 76.2 tok/s67.0–85.3calibrated estimate ±12% |
| 64k | 8.4 GBRuns great est. 58.7 tok/s51.6–65.7calibrated estimate ±12% | 7.4 GBRuns great est. 67.0 tok/s59.0–75.1calibrated estimate ±12% | 6.8 GBRuns great est. 72.6 tok/s63.9–81.3calibrated estimate ±12% |
| 128k | 11.1 GBRuns great est. 46.3 tok/s40.7–51.8calibrated estimate ±12% | 9.1 GBRuns great est. 57.6 tok/s50.7–64.5calibrated estimate ±12% | 8.0 GBRuns great est. 66.3 tok/s58.4–74.3calibrated estimate ±12% |
| 256k | 14.5 GBRuns slowly est. 3.1 tok/s2.5–3.7calibrated estimate ±20% | 12.6 GBRuns slowly est. 10.7 tok/s8.6–12.9calibrated estimate ±20% | 10.5 GBRuns great est. 56.6 tok/s49.8–63.3calibrated estimate ±12% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs great up to 128k context with an f16 KV cache, and up to 128k with q8_0.
Estimated speed
- Generation
- est. 76.6 tok/s (67.4–85.8, calibrated estimate ±12%)
- Prompt processing
- about 3,375 tok/s (2,025–4,724, rough estimate ±40%)
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 5070 | Q8_0 | llama.cpp | 32k | 3,001.0 | 67.3 | — | localmaxxing.com | 2026-09-10 |
If this is not enough
A larger model that also runs well on the GeForce RTX 5070
Runs wellQwen3.5 35B-A3B: Runs well, est. 63.3 tok/s (44.3–82.3, theoretical estimate ±30%)
A larger quant that still runs great
Runs greatQ8_0 (9.79 GB): Runs great, 67.3 tok/s measured at 32k context (1 run; 8k estimate 46.8 tok/s, 41.2–52.4, calibrated estimate ±12%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Qwen3.5-9B-Q4_K_M.gguf (5.87 GB) from unsloth/Qwen3.5-9B-MTP-GGUF.
llama-server -hf unsloth/Qwen3.5-9B-MTP-GGUF:Q4_K_M -c 8192 -ngl all -fa on- -ngl all loads every layer on the GPU.
- -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 0.14 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/unsloth/Qwen3.5-9B-MTP-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.
/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
Qwen3.5 9B on other hardware
Frequently asked questions
Can I run Qwen3.5 9B on the GeForce RTX 5070?
Q4_K_M fits in VRAM with 5.4 GB to spare: Runs great, est. 76.6 tok/s (67.4–85.8, calibrated estimate ±12%). Q4_K_M needs 6.0 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 Qwen3.5 9B use on the GeForce RTX 5070?
At Q4_K_M the verdict stays Runs great up to 128k context with an f16 KV cache (4.3 GB of KV) and up to 128k with a q8_0 KV cache (2.3 GB). The model supports up to 256k.
Which quant should I use, and how fast is it?
Q4_K_M (5.87 GB) is the recommended balance: Runs great, est. 76.6 tok/s (67.4–85.8, calibrated estimate ±12%). The largest tracked file that still gets Runs great is Q8_0 (9.79 GB), 67.3 tok/s measured at 32k context (1 run; 8k estimate 46.8 tok/s, 41.2–52.4, calibrated estimate ±12%).
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.