What LLMs can four RTX 3090 GPUs run?
Four RTX 3090 GPUs (96 GB in all) run 48 of the 69 open models tracked here at Q4_K_M with 8K context; the largest is Mistral Medium 3.5 128B (127.7B parameters, 84.2 GB). They read memory at 3,744 GB/s in total. 2 more fit at a lower precision. Each fit leaves at least 0.5 GB free on 4× RTX 3090 (why a card does not give a model all its memory). The speeds assume tensor parallelism, as in vLLM or SGLang; llama.cpp splits layers across cards by default and then writes at about the speed of one card. How the model is divided between the cards, and what each one holds.
Models that fit on 4× RTX 3090
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. Memory is for the whole group: weights, cache and 10% split across the 4 cards, plus 0.5 GB of runtime on each card (2 GB in all). gpt-oss-120b (69.2 GB, 158–334 tokens/s on 4× RTX 3090) and gpt-oss-20b (16.3 GB, 172–373 tokens/s on 4× RTX 3090) are listed at MXFP4 only, as published.
| Model | Best precision | Memory | Longest context | Tokens/s |
|---|---|---|---|---|
| Step 3.7 Flash 196B-A11B 201.4B, 11B active | GGUF Q2_K | 88.9 GB | 135K | 124–247 |
| Qwen3.8 Flash Next 180.0B, 6B active | GGUF Q3_K_M | 92.3 GB | 130K | 159–336 |
| Mistral Medium 3.5 128B 127.7B | GGUF Q4_K_M | 84.2 GB | 37K | 24–33 |
| Qwen3.5 122B-A10B 125.1B, 10B active | GGUF Q5_K_M | 93.0 GB | 103K | 100–193 |
| Nemotron 3.5 Super VL 120B-A12B 124.5B, 12B active | GGUF Q5_K_M | 92.4 GB | 364K | 90–170 |
| Nemotron 3 Super 120B-A12B 123.6B, 12B active | GGUF Q5_K_M | 91.8 GB | 256K (full) | 90–170 |
| gpt-oss-120b 116.8B, 5.1B active | As published (MXFP4) | 69.2 GB | 128K (full) | 158–334 |
| AliceAI Foundation 80B-A3B 81.3B, 3B active | GGUF Q8_0 | 90.7 GB | 194K | 153–322 |
| Qwen3-Coder-Next 79.7B, 3B active | GGUF Q8_0 | 88.9 GB | 256K (full) | 153–322 |
| Kolibri-1 78.1B, 3.5B active | GGUF Q8_0 | 87.3 GB | 256K (full) | 143–293 |
| Llama 3.1 70B 70.6B | GGUF Q8_0 | 81.5 GB | 48K | 24–35 |
| K2-Horizon MoVA 36B-A4B 37.4B, 4B active | FP16 / BF16 | 80.4 GB | 81K | 83–155 |
| Ornith 1.5 35B-A3B 36.0B, 3B active | FP16 / BF16 | 75.8 GB | 256K (full) | 111–218 |
| Qwen3.6 35B-A3B 36.0B, 3B active | FP16 / BF16 | 75.8 GB | 256K (full) | 111–218 |
| Ornith 1.0 35B 35.1B, 3B active | FP16 / BF16 | 74.1 GB | 256K (full) | 111–218 |
| LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active | FP16 / BF16 | 68.4 GB | 64K (full) | 93–176 |
| Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active | FP16 / BF16 | 66.8 GB | 256K (full) | 102–197 |
| Gemma 4 31B 31.3B | FP16 / BF16 | 68.1 GB | 256K (full) | 29–41 |
| GLM-4.7 Flash 31.2B, 3B active | FP16 / BF16 | 66.4 GB | 198K (full) | 108–211 |
| Xing 4.0 29B-A4B 31.2B, 4B active | FP16 / BF16 | 66.3 GB | 256K (full) | 91–173 |
| Qwen3-Coder 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 65.4 GB | 256K (full) | 99–190 |
| Qwen3 30B-A3B 30.5B, 3.3B active | FP16 / BF16 | 65.4 GB | 40K (full) | 99–190 |
| Muse Glimmer 30B 29.8B | FP16 / BF16 | 63.2 GB | 128K (full) | 31–44 |
| Granite 4.2 30B 29.3B | FP16 / BF16 | 64.2 GB | 121K | 30–44 |
| ThinkingCap Qwen3.8 27B 27.8B | FP16 / BF16 | 59.5 GB | 256K (full) | 33–47 |
| Qwen3.6 27B 27.8B | FP16 / BF16 | 59.5 GB | 256K (full) | 33–47 |
| Qwen3.8 27B 27.8B | FP16 / BF16 | 59.5 GB | 256K (full) | 33–47 |
| Hemmingway-1 27B 27.3B | FP16 / BF16 | 58.5 GB | 256K (full) | 33–48 |
| Gemma 4 26B-A4B 25.8B, 4B active | FP16 / BF16 | 55.4 GB | 256K (full) | 90–171 |
| gpt-oss-20b 20.9B, 3.6B active | As published (MXFP4) | 16.3 GB | 128K (full) | 172–373 |
| Gemma 4 12B 12.0B | FP16 / BF16 | 27.2 GB | 256K (full) | 65–100 |
| ZDTaichu 5.0 9B 9.8B | FP16 / BF16 | 22.3 GB | 128K (full) | 76–119 |
| Ornith 1.5 9B 9.7B | FP16 / BF16 | 22.1 GB | 256K (full) | 77–121 |
| Qwen3.5 9B 9.7B | FP16 / BF16 | 22.1 GB | 256K (full) | 77–121 |
| LensVLM 9B 9.4B | FP16 / BF16 | 21.6 GB | 256K (full) | 78–123 |
| Ornith 1.0 9B 9.4B | FP16 / BF16 | 21.6 GB | 256K (full) | 78–123 |
| MiMo V2.6 Distill Qwen 9B 9.4B | FP16 / BF16 | 21.6 GB | 256K (full) | 78–123 |
| Bespoke Nimble 9B 9.0B | FP16 / BF16 | 20.6 GB | 8K (full) | 81–129 |
| Granite 4.2 8B 8.8B | FP16 / BF16 | 21.4 GB | 128K (full) | 79–124 |
| LFM2.5 8B-A1B 8.5B, 1.5B active | FP16 / BF16 | 19.5 GB | 128,000 (full) | 160–338 |
| Qwen3 8B 8.2B | FP16 / BF16 | 20.0 GB | 40K (full) | 83–132 |
| Llama 3.1 8B 8.0B | FP16 / BF16 | 19.6 GB | 128K (full) | 85–135 |
| Gemma 4 E4B 8.0B | FP16 / BF16 | 18.6 GB | 128K (full) | 88–142 |
| Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active | FP16 / BF16 | 18.2 GB | 128K (full) | 170–368 |
| LightOnOCR-3 4B 4.5B | FP16 / BF16 | 11.6 GB | 256K (full) | 124–216 |
| Spark-X2.5 4B 4.1B | FP16 / BF16 | 10.9 GB | 1M (full) | 130–228 |
| Nemotron 3 Nano 4B 4.0B | FP16 / BF16 | 10.3 GB | 256K (full) | 135–239 |
| Granite 4.2 3B 3.7B | FP16 / BF16 | 10.2 GB | 128K (full) | 135–241 |
| MiniCPM5 2B 2.5B | FP16 / BF16 | 7.52 GB | 128K (full) | 164–311 |
| Limite 1B Violetto 1.0B | FP16 / BF16 | 4.29 GB | 128K (full) | 218–477 |
Too big for 4× RTX 3090 at 8K context
How many RTX 3090 cards these need at Q4_K_M with 8K context and 0.5 GB to spare, in one tensor-parallel group; some need more than 8.
- MiniMax M2.7: 8 cards
- DeepSeek V4 Flash: 8 cards
- DeepSeek V4 Flash 0731: more than 8 cards
- MiMo V2.6 Flash: more than 8 cards
- IQuest-Q1 320B-A15B: more than 8 cards
- GLM-5.3 Flash: more than 8 cards
- MiniMax M3: more than 8 cards
- Ling 3.1 Flash 560B-A25B: more than 8 cards
- DeepSeek V3 / R1: more than 8 cards
- DeepSeek V3.2: more than 8 cards
- GLM-5.3: more than 8 cards
- GLM-5.2: more than 8 cards
- DeepSeek V4.1 Flash 1017: more than 8 cards
- DeepSeek V4.1 Flash: more than 8 cards
- Hy4 Preview 770B-A49B: more than 8 cards
- MiMo V2.6 Pro: more than 8 cards
- DeepSeek V4 Pro: more than 8 cards
- Qwen3.8 2.4T-A95B: more than 8 cards
- Kimi K3: more than 8 cards
Best models for 4× RTX 3090 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 group takes at that precision and the writing speed with that context in the cache, from 3,744 GB/s of bandwidth. Memory is for the whole group: weights, cache and 10% split across the 4 cards, plus 0.5 GB of runtime on each card (2 GB in all). Memory and left over add up to the group's 96 GB. gpt-oss-120b is at MXFP4, as published.
| Model | Context | Memory | Left over | Longest context | Tokens/s |
|---|---|---|---|---|---|
| Mistral Medium 3.5 128B 127.7B | 32K | 93.3 GB | 2.75 GB | 37K | 21–30 |
| Qwen3.5 122B-A10B 125.1B | 32K | 80.4 GB | 15.6 GB | 256K (full) | 104–202 |
| Nemotron 3.5 Super VL 120B-A12B 124.5B | 32K | 79.4 GB | 16.6 GB | 1M (full) | 98–188 |
| Nemotron 3 Super 120B-A12B 123.6B | 32K | 78.9 GB | 17.1 GB | 256K (full) | 98–188 |
| gpt-oss-120b 116.8B, MXFP4 | 32K | 70.1 GB | 25.9 GB | 128K (full) | 140–287 |
llama-server commands for 4× RTX 3090
GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on 4× RTX 3090. No -ngl: llama.cpp’s -fit, on by default since b7440, places the layers.
Mistral Medium 3.5 128B, Q4_K_M with 37K tokens: 21–30 tokens/s on 4× RTX 3090
llama-server -hf unsloth/Mistral-Medium-3.5-128B-GGUF:Q4_K_M -c 37888Q4_K_M/
Qwen3.5 122B-A10B, Q4_K_M with 256K tokens: 68–125 tokens/s on 4× RTX 3090
llama-server -hf unsloth/Qwen3.5-122B-A10B-GGUF:Q4_K_M -c 262144Q4_K_M/
How fast 4× RTX 3090 writes as the context fills
Tokens per second at Q4_K_M (gpt-oss-120b at MXFP4) for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the group 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 3,744 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, Ryzen AI Max+ 395 (128 GB), 256 GB/s, writes 93% slower by bandwidth; M4 Max Mac (128 GB), 546 GB/s, writes 85% slower by bandwidth; RTX PRO 6000 Blackwell, 1,792 GB/s, writes 52% slower by bandwidth.
| Model | 8K | 32K | Longest | At the longest | Q8_0 at 8K | Bandwidth ceiling at 8K | 1,000-token reply at 32K |
|---|---|---|---|---|---|---|---|
| Mistral Medium 3.5 128B | 24–33 | 21–30 | 37K | 21–30 | Does not fit | 47 | 33–47 s |
| Qwen3.5 122B-A10B | 110–216 | 104–202 | 256K | 68–125 | Does not fit | 599 | 5–10 s |
| Nemotron 3.5 Super VL 120B-A12B | 100–192 | 98–188 | 1M | 57–102 | Does not fit | 511 | 5–10 s |
| Nemotron 3 Super 120B-A12B | 100–192 | 98–188 | 256K | 84–158 | Does not fit | 511 | 5–10 s |
| gpt-oss-120b | 158–334 | 140–287 | 128K | 96–185 | MXFP4 only | 1,178 | 3–7 s |
| AliceAI Foundation 80B-A3B | 189–421 | 171–370 | 256K | 92–175 | 153–322 | 1,857 | 3–6 s |
| Qwen3-Coder-Next | 189–421 | 171–370 | 256K | 92–175 | 153–322 | 1,857 | 3–6 s |
| Kolibri-1 | 178–390 | 165–353 | 256K | 98–187 | 143–293 | 1,583 | 3–6 s |
| Llama 3.1 70B | 39–57 | 34–49 | 128K | 22–31 | 24–35 | 83 | 20–29 s |
| K2-Horizon MoVA 36B-A4B | 141–290 | 88–166 | 340K | 15–26 | 115–226 | 929 | 6–11 s |
Near misses on 4× RTX 3090
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 |
|---|---|---|
| Qwen3.8 Flash Next 180.0B | 18.4 GB | GGUF Q3_K_M with 32K, 93.0 GB |
| Step 3.7 Flash 196B-A11B 201.4B | 32.6 GB | GGUF IQ3_XXS with 32K, 88.9 GB |
What more cards or MoE offload unlock
8× RTX 3090 (192 GB) would add 4 models at Q4_K_M with 8K context: Qwen3.8 Flash Next, Step 3.7 Flash 196B-A11B, MiniMax M2.7, DeepSeek V4 Flash.
Run locally or rent four RTX 3090 GPUs?
Renting four RTX 3090 GPUs costs a median of $1.08 an hour on demand ($0.27 per card), across the providers getdeploying.com tracks (checked ), so every $1,000 of purchase price equals about 926 hours of rent, 39 days around the clock. At that rate Mistral Medium 3.5 128B at Q4_K_M with 32K tokens, the biggest model here that fits with a long context, writes 21–30 tokens/s for one request: about $9.9–14 per million generated tokens. Electricity and resale value are left out, and no card prices are listed because 2026 street prices vary widely between sellers; compare current ones before you decide.
Serving with vLLM on 4× RTX 3090
The KV cache pool vLLM builds with --tensor-parallel-size 4 at its default --gpu-memory-utilization 0.92 and --max-model-len 32768, weights as published, from the vLLM KV cache calculator. The pool is shared by every request; the last column assumes 4,096 tokens each.
| Model | KV cache per GPU | Pool (tokens) | At 32K | Requests at 4K |
|---|---|---|---|---|
| gpt-oss-120b | 4.1 GB | 420,033 | 12.82x | 57 |
| Qwen3 30B-A3B | 6.0 GB | 260,288 | 7.94x | 63 |
| gpt-oss-20b | 16.8 GB | 2,596,364 | 79.23x | 357 |
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 on each card 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 48 here.
| Setup | How it relates | Memory | Bandwidth | Models at Q4_K_M |
|---|---|---|---|---|
| Ryzen AI Max+ 395 (128 GB) | Same memory | 96 GB usable | 256 GB/s | 48 (same) |
| M4 Max Mac (128 GB) | Same memory | 96 GB usable | 546 GB/s | 48 (same) |
| RTX PRO 6000 Blackwell | Same memory | 96 GB | 1,792 GB/s | 48 (same) |
| H100 SXM | Next size down | 80 GB | 3,350 GB/s | 47 (−1) |
| DGX Spark (128 GB) | Next size up | 120 GB usable | 273 GB/s | 49 (+1) |
| RTX 3090 | Fewer of this card | 24 GB | 936 GB/s | 38 (−10) |
| 2× RTX 3090 | Fewer of this card | 48 GB | 1,872 GB/s | 40 (−8) |
Model numbers read from Hugging Face on .