What LLMs can a CPU-only PC with 32 GB of dual-channel DDR4-3200 run?
A CPU-only PC with 32 GB of dual-channel DDR4-3200 gives a model 26 GB usable of memory (system RAM minus about 6 GB for the OS and apps) and 51.2 GB/s of bandwidth. Of the 61 open models tracked here, 35 fit at Q4_K_M with an 8,192-token context, each leaving at least 0.5 GB free on this PC.
With no GPU, llama.cpp (or Ollama and LM Studio on top of it) keeps the whole model in the 32 GB of DDR4-3200 and reads all of its active weights over the memory bus for every token. A model fits here when weights, cache and overhead stay within the RAM minus 6 GB held back for the operating system and the apps next to it; close the browser and that reserve shrinks. MoE offload does not apply, since everything already runs from RAM, but mixture-of-experts models still help most: only their active experts are read per token. The speeds use 50–80% of the rated 51.2 GB/s for dense models and 30–55% for MoE models, from public CPU-only llama.cpp runs. Reading the prompt is compute-bound on a CPU, often tens of tokens per second, so a long prompt can take minutes before the first word; a GPU of any size speeds that part up even when the weights stay in RAM.
Models that fit in 32 GB of DDR4-3200 (dual channel, CPU only)
The most precise weights that still fit with 8K tokens of context and one request, the memory that takes, the longest context at that precision and the writing speed. Bigger models come first. gpt-oss-20b (14.8 GB, 5.9–11 tokens/s on this PC) is listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | GGUF Q4_K_M | 25.4 GB | 8K | 3.8–7.0 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | GGUF Q4_K_M | 23.0 GB | 126K | 7.7–14 |
| Qwen3.6 35B-A3B 36.0B, 3B active | GGUF Q4_K_M | 23.0 GB | 126K | 7.7–14 |
| Ornith 1.0 35B 35.1B, 3B active | GGUF Q4_K_M | 22.4 GB | 150K | 7.7–14 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | GGUF Q5_K_M | 24.4 GB | 24K | 4.7–8.6 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | GGUF Q5_K_M | 23.5 GB | 256K (full) | 6.0–11 |
| Gemma 4 31B 31.3B | GGUF Q5_K_M | 25.2 GB | 11K | 1.1–1.7 |
| GLM-4.7 Flash 31.2B, 3B active | GGUF Q5_K_M | 23.6 GB | 41K | 5.9–11 |
| Xing 4.0 29B-A4B 31.2B, 4B active | GGUF Q5_K_M | 23.6 GB | 48K | 4.7–8.7 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | GGUF Q5_K_M | 23.5 GB | 27K | 4.8–8.9 |
| Qwen3 30B-A3B 30.5B, 3.3B active | GGUF Q5_K_M | 23.5 GB | 27K | 4.8–8.9 |
| Muse Glimmer 30B 29.8B | GGUF Q5_K_M | 22.3 GB | 128K (full) | 1.2–1.9 |
| Granite 4.2 30B 29.3B | GGUF Q5_K_M | 24.0 GB | 13K | 1.1–1.8 |
| Qwen3.6 27B 27.8B | GGUF Q6_K | 24.4 GB | 24K | 1.1–1.8 |
| Qwen3.8 27B 27.8B | GGUF Q6_K | 24.4 GB | 24K | 1.1–1.8 |
| Hemmingway-1 27B 27.3B | GGUF Q6_K | 24.0 GB | 29K | 1.1–1.8 |
| Gemma 4 26B-A4B 25.8B, 4B active | GGUF Q6_K | 22.7 GB | 139K | 4.1–7.5 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 14.8 GB | 128K (full) | 5.9–11 |
| Gemma 4 12B 12.0B | GGUF Q8_0 | 14.2 GB | 256K (full) | 1.9–3.1 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 20.8 GB | 128K (full) | 1.3–2.1 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 151K | 1.3–2.1 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 20.6 GB | 151K | 1.3–2.1 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 20.1 GB | 166K | 1.3–2.1 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 20.1 GB | 166K | 1.3–2.1 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 19.9 GB | 40K | 1.3–2.2 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 18.0 GB | 128,000 (full) | 4.9–9.0 |
| Qwen3 8B 8.2B | FP16 / BF16 | 18.5 GB | 40K (full) | 1.5–2.3 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 18.1 GB | 62K | 1.5–2.4 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 17.1 GB | 128K (full) | 1.6–2.5 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 16.7 GB | 128K (full) | 5.7–11 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 9.35 GB | 425K | 3.0–4.7 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 8.78 GB | 256K (full) | 3.2–5.1 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 8.69 GB | 128K (full) | 3.2–5.1 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 6.02 GB | 128K (full) | 4.7–7.6 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 2.79 GB | 128K (full) | 11–18 |
What one GPU adds to this PC
With a graphics card, llama.cpp's --n-cpu-moe keeps the routed experts of the first layers in this PC's
32 GB of DDR4-3200 and everything else on the card, so a mixture-of-experts model can run far bigger than the card and far faster than
the CPU alone. Each measured GGUF file below at its own 32K context (or its whole window): tokens per second on the CPU alone
at MXFP4 where it fits, then the smallest --n-cpu-moe and tokens per second with an
RTX 5060 Ti 16GB or an RTX 4090, the experts read at 51.2 GB/s.
"More RAM" means the experts left on the CPU need more than the 26 GB this PC leaves for a model.
| Model | GGUF | CPU only | --n-cpu-moe, RTX 5060 Ti 16GB | --n-cpu-moe, RTX 4090 | Tokens/s, RTX 5060 Ti 16GB | Tokens/s, RTX 4090 |
|---|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | 4.8–8.8 | 0 | 0 | 37–64 | 78–138 |
| Gemma 4 26B-A4B | UD-Q4_K_M | 4.5–8.2 | 3 | 0 | 29–50 | 73–128 |
| Qwen3 30B-A3B | Q4_K_M | 2.9–5.4 | 15 | 0 | 17–28 | 54–93 |
| GLM-4.7 Flash | Q4_K_M | 4.3–7.8 | 12 | 0 | 21–36 | 65–114 |
| Nemotron 3 Nano 30B-A3B | Q4_K_M | 6.6–12 | 21 | 0 | 21–36 | 94–167 |
| Ornith 1.5 35B-A3B | Q4_K_M | 6.1–11 | 13 | 0 | 30–52 | 96–171 |
| Qwen3.6 35B-A3B | UD-Q4_K_M | 6.1–11 | 13 | 0 | 27–45 | 81–142 |
| Qwen3-Coder-Next | Q4_K_M | Does not fit | 35 (more RAM) | 26 | — | 23–40 |
| gpt-oss-120b | MXFP4 | Does not fit | 29 (more RAM) | 24 (more RAM) | — | — |
| Qwen3.5 122B-A10B | Q4_K_M | Does not fit | 42 (more RAM) | 36 (more RAM) | — | — |
| Qwen3.8 Flash Next | UD-Q4_K_XL | Does not fit | 42 (more RAM) | 37 (more RAM) | — | — |
Too big for 32 GB of DDR4-3200 (dual channel, CPU only) at 8K context
How much memory each of these lacks at Q4_K_M (gpt-oss-120b at MXFP4) with 8K context, one request and 0.5 GB kept free, out of the 26 GB this PC leaves for a model; more RAM or a smaller quant is the way in.
- Llama 3.1 70B: 21.5 GB short
- Qwen3-Coder-Next: 24.6 GB short
- AliceAI Foundation 80B-A3B: 25.6 GB short
- gpt-oss-120b: 42.2 GB short
- Nemotron 3 Super 120B-A12B: 51.7 GB short
- Qwen3.5 122B-A10B: 52.7 GB short
- Mistral Medium 3.5 128B: 57.2 GB short
- Qwen3.8 Flash Next: 86.8 GB short
- Step 3.7 Flash 196B-A11B: 100 GB short
- MiniMax M2.7: 119 GB short
- DeepSeek V4 Flash: 156 GB short
- DeepSeek V4 Flash 0731: 164 GB short
- MiMo V2.6 Flash: 168 GB short
- IQuest-Q1 320B-A15B: 176 GB short
- GLM-5.3 Flash: 174 GB short
- MiniMax M3: 241 GB short
- DeepSeek V3 / R1: 400 GB short
- DeepSeek V3.2: 400 GB short
- GLM-5.3: 443 GB short
- GLM-5.2: 443 GB short
- DeepSeek V4.1 Flash: 449 GB short
- Hy4 Preview 770B-A49B: 459 GB short
- MiMo V2.6 Pro: 610 GB short
- DeepSeek V4 Pro: 967 GB short
- Qwen3.8 2.4T-A95B: 1,492 GB short
- Kimi K3: 1,698 GB short
Best models for 32 GB of DDR4-3200 (dual channel, CPU only) with room for long context
The biggest models that still fit at Q4_K_M with 32,768 tokens of context (or their whole window, if shorter), the memory left over, the longest context the PC takes at that precision and the writing speed with that context in the cache, from 51.2 GB/s of bandwidth.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| Ornith 1.5 35B-A3B 36.0B | 32K | 23.5 GB | 2.53 GB | 126K | 6.1–11 |
| Qwen3.6 35B-A3B 36.0B | 32K | 23.5 GB | 2.53 GB | 126K | 6.1–11 |
| Ornith 1.0 35B 35.1B | 32K | 22.9 GB | 3.05 GB | 150K | 6.1–11 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B | 32K | 22.6 GB | 3.38 GB | 64K (full) | 3.4–6.3 |
| Nemotron 3 Nano 30B-A3B 31.6B | 32K | 20.3 GB | 5.72 GB | 256K (full) | 6.6–12 |
llama-server commands for a CPU-only PC with 32 GB of dual-channel DDR4-3200
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on this PC. -ngl 0 keeps every layer in system RAM, even in a CUDA or Vulkan build of llama.cpp.
Ornith 1.5 35B-A3B, Q4_K_M with 121K tokens: 3.5–6.4 tokens/s on this PC
llama-server -hf bartowski/Ornith-1.5-35B-A3B-GGUF:Q4_K_M -c 123904 -ngl 0Ornith-1.5-35B-A3B-Q4_K_M.gguf, 21.9 GB: 25.5 GB used and 514 MB free of this PC's 26 GB usable, 3.5–6.4 tokens/s once the 121K cache is full.
Qwen3.6 35B-A3B, Q4_K_M with 126K tokens: 3.4–6.3 tokens/s on this PC
llama-server -hf ggml-org/Qwen3.6-35B-A3B-GGUF:Q4_K_M -c 129024 -ngl 0Qwen3.6-35B-A3B-Q4_K_M.gguf, 20.4 GB: 25.5 GB used and 522 MB free of this PC's 26 GB usable, 3.4–6.3 tokens/s once the 126K cache is full.
Ornith 1.0 35B, Q4_K_M with 143K tokens: 3.2–5.8 tokens/s on this PC
llama-server -hf bartowski/deepreinforce-ai_Ornith-1.0-35B-GGUF:Q4_K_M -c 146432 -ngl 0deepreinforce-ai_Ornith-1.0-35B-Q4_K_M.gguf, 21.4 GB: 25.5 GB used and 525 MB free of this PC's 26 GB usable, 3.2–5.8 tokens/s once the 143K cache is full.
How fast this PC writes as the context fills
Tokens per second at Q4_K_M for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the PC holds, and at Q8_0 with 8K where that fits. Each token reads the active weights and the whole cache once, so the speed follows the 51.2 GB/s of bandwidth and falls as the cache grows. The ceiling is that bandwidth divided by the bytes read per token, which no runtime reaches; the reply time is for 1,000 new tokens with 32K (or the longest context) in the cache, prompt processing aside. With the same memory, CPU only, DDR5-5600 dual channel (32 GB), 89.6 GB/s, writes 74% faster on average.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| Ornith 1.5 35B-A3B | 7.7–14 | 6.1–11 | 126K | 3.4–6.3 | Does not fit | 26 | 89–163 s |
| Qwen3.6 35B-A3B | 7.7–14 | 6.1–11 | 126K | 3.4–6.3 | Does not fit | 26 | 89–163 s |
| Ornith 1.0 35B | 7.7–14 | 6.1–11 | 150K | 3.1–5.7 | Does not fit | 26 | 89–163 s |
| LLM-jp-4.1 32B-A3B Thinking | 5.3–9.8 | 3.4–6.3 | 64K | 2.3–4.3 | Does not fit | 18 | 159–292 s |
| Nemotron 3 Nano 30B-A3B | 7.0–13 | 6.6–12 | 256K | 4.1–7.5 | Does not fit | 24 | 83–152 s |
| Gemma 4 31B | 1.2–2.0 | 1.1–1.8 | 50K | 1.0–1.7 | Does not fit | 2 | 559–895 s |
| GLM-4.7 Flash | 6.7–12 | 4.3–7.8 | 99K | 2.1–3.8 | Does not fit | 23 | 128–235 s |
| Xing 4.0 29B-A4B | 5.4–10 | 3.9–7.1 | 116K | 1.9–3.6 | Does not fit | 18 | 140–257 s |
| Qwen3-Coder 30B-A3B | 5.4–10 | 2.9–5.4 | 58K | 2.0–3.6 | Does not fit | 18 | 186–341 s |
| Qwen3 30B-A3B | 5.4–10 | 2.9–5.4 | 40K | 2.5–4.7 | Does not fit | 18 | 186–341 s |
32 GB of DDR4-3200 (dual channel, CPU only) against the other CPU-only PCs with 26 GB usable
The same models fit on every CPU-only PC with 26 GB usable, so speed is what separates them. By bandwidth the CPU only, DDR4-3200 dual channel (32 GB) is the slowest of the 2 CPU-only PCs with 26 GB usable at 51.2 GB/s. Over the 12 biggest models that fit at Q4_K_M with 8K context, the CPU only, DDR5-5600 dual channel (32 GB), at 89.6 GB/s, writes 74% faster than the CPU only, DDR4-3200 dual channel (32 GB). Each cell below is the other CPU-only PC's tokens per second minus this one's, midpoints of the estimated ranges.
| Model | CPU only, DDR4-3200 dual channel (32 GB) tokens/s | CPU only, DDR5-5600 dual channel (32 GB) |
|---|---|---|
| K2-Horizon MoVA 36B-A4B | 3.8–7.0 | +4.0 |
| Ornith 1.5 35B-A3B | 7.7–14 | +8.0 |
| Qwen3.6 35B-A3B | 7.7–14 | +8.0 |
| Ornith 1.0 35B | 7.7–14 | +8.0 |
| LLM-jp-4.1 32B-A3B Thinking | 5.3–9.8 | +5.6 |
| Nemotron 3 Nano 30B-A3B | 7.0–13 | +7.4 |
| Gemma 4 31B | 1.2–2.0 | +1.2 |
| GLM-4.7 Flash | 6.7–12 | +7.1 |
| Xing 4.0 29B-A4B | 5.4–10 | +5.7 |
| Qwen3-Coder 30B-A3B | 5.4–10 | +5.7 |
| Qwen3 30B-A3B | 5.4–10 | +5.7 |
| Muse Glimmer 30B | 1.4–2.2 | +1.4 |
Near misses on 32 GB of DDR4-3200 (dual channel, CPU only)
Models that miss at Q4_K_M with 32,768 tokens of context (or their whole window), or leave under 0.5 GB free there, but fit with a shorter context or 3-bit weights, or are short by at most a quarter of the memory and fit with MoE offload. Smallest shortfall first.
| Model | Short by at Q4_K_M | Fits instead |
|---|---|---|
| Granite 4.2 30B 29.3B | 1.45 GB | Q4_K_M with 24K context |
| K2-Horizon MoVA 36B-A4B 37.4B | 4.31 GB | Q4_K_M with 8K context |
How these numbers are worked out
Memory is the weights at each precision plus the KV cache for 8,192 tokens in FP16 and runtime overhead (0.5 GB plus 10%), computed from each model's files on Hugging Face with the LLM VRAM Calculator. Speed is estimated from memory bandwidth and the parameters read per token with the LLM Speed Calculator. Both tools take any other model from Hugging Face.
Nearby GPUs and setups
How many of the tracked models each one holds at Q4_K_M with 8K context, against 35 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| CPU only, DDR5-5600 dual channel (32 GB) | Same memory | 26 GB usable | 89.6 GB/s | 35 (same) |
| CPU only, DDR4-3200 dual channel (16 GB) | Next size down | 10 GB usable | 51.2 GB/s | 17 (−18) |
| CPU only, DDR4-3200 dual channel (64 GB) | Next size up | 58 GB usable | 51.2 GB/s | 38 (+3) |
Model numbers read from Hugging Face on .