What LLMs can Desktop PC · 2× GeForce RTX 3090 (48 GB) · 128 GB RAM run?
48 GB VRAM + 128 GB RAM, 936 GB/s memory bandwidth, about $2,884. Every open model in our catalog, rated by the engine for chat and for agentic coding.
Runs well (35)
Fast enough and with memory to spare for everyday chat.
| Model | Chat | Agentic coding | Quant | Chat speed |
|---|---|---|---|---|
| Qwen3.6 35B-A3BAlibaba QwenOpen source | Excellent | Excellent | Q6_K | 140–230 tok/s |
| Gemma 4 26B-A4BGoogleOpen source | Excellent | Comfortable | Q6_K | 110–180 tok/s |
| Qwen3.5 35B-A3BAlibaba QwenOpen source | Excellent | Comfortable | Q6_K | 140–230 tok/s |
| LFM2.5 24B-A2BLiquid AI | Excellent | Comfortable | Q4_K_M | 270–450 tok/s |
| gpt-oss-20bOpenAIOpen source | Excellent | Comfortable | MXFP4 | 160–270 tok/s |
| GLM-4.7-Flash (30B-A3B)Z.aiOpen source | Excellent | Comfortable | Q6_K | 130–220 tok/s |
| Qwen3 30B-A3B Instruct 2507Alibaba QwenOpen source | Excellent | Comfortable | Q6_K | 120–190 tok/s |
| Qwen3-Next 80B-A3B InstructAlibaba QwenOpen source | Excellent | Acceptable | Q3_K_M | 230–380 tok/s |
| Kolibri-1 (78B-A3.5B)Aleph AlphaOpen source | Excellent | Borderline | Q4_K_M | 160–270 tok/s |
| Qwen3.8 27BAlibaba QwenOpen source | Excellent | Comfortable | Q5_K_M | 25–40 tok/s |
| Qwen3-Coder 30B-A3B InstructAlibaba QwenOpen source | Excellent | Comfortable | Q6_K | 120–190 tok/s |
| Qwen3.6 27BAlibaba QwenOpen source | Excellent | Comfortable | Q5_K_M | 25–40 tok/s |
| MiniCPM5 2BOpenBMBOpen source | Excellent | Acceptable | Q8_0 | 170–280 tok/s |
| Qwen3.5 9BAlibaba QwenOpen source | Excellent | Acceptable | Q6_K | 55–95 tok/s |
| Gemma 4 31BGoogleOpen source | Excellent | Comfortable | Q4_K_M | 25–40 tok/s |
| Qwen3.5 27BAlibaba QwenOpen source | Excellent | Comfortable | Q5_K_M | 25–40 tok/s |
| Qwen3.5 4BAlibaba QwenOpen source | Excellent | Acceptable | Q8_0 | 100–170 tok/s |
| Gemma 4 12BGoogleOpen source | Excellent | Acceptable | Q6_K | 45–75 tok/s |
| Nemotron 3.5 Bolt 74B-A9BNVIDIA | Excellent | Acceptable | 4-bit | 65–110 tok/s |
| GLM-4.6V-Flash (10B)Z.aiOpen source | Excellent | Acceptable | Q6_K | 55–90 tok/s |
| Qwen3 8BAlibaba QwenOpen source | Excellent | Acceptable | Q6_K | 65–110 tok/s |
| Qwen3 14BAlibaba QwenOpen source | Excellent | Acceptable | Q6_K | 35–60 tok/s |
| DeepSeek-R1-0528-Qwen3-8BDeepSeekOpen source | Excellent | Acceptable | Q6_K | 65–110 tok/s |
| Muse Glimmer (30B)MetaOpen source | Excellent | Acceptable | Q4_K_M | 30–45 tok/s |
| Gemma 4 E4BGoogleOpen source | Excellent | Borderline | Q6_K | 70–110 tok/s |
| Mistral Small 3.2 (24B)Mistral AIOpen source | Excellent | Acceptable | Q5_K_M | 25–45 tok/s |
| Phi-4 (14B)MicrosoftOpen source | Excellent | Acceptable | Q6_K | 35–60 tok/s |
| Llama 3.2 3B InstructMetaOpen weights (restricted) | Excellent | Technically Runs | Q6_K | 150–260 tok/s |
| Llama 3.1 8B InstructMetaOpen weights (restricted) | Excellent | Borderline | Q6_K | 65–110 tok/s |
| Gemma 3 27BGoogleOpen weights (restricted) | Excellent | Acceptable | Q5_K_M | 25–40 tok/s |
| Devstral Small 2 (24B)Mistral AIOpen source | Excellent | Comfortable | Q5_K_M | 25–45 tok/s |
| Qwen3 32BAlibaba QwenOpen source | Excellent | Acceptable | Q5_K_M | 19–30 tok/s |
| Qwen2.5-Coder 7B InstructAlibaba QwenOpen source | Excellent | Acceptable | Q6_K | 70–120 tok/s |
| Qwen2.5-Coder 32B InstructAlibaba QwenOpen source | Excellent | Acceptable | Q5_K_M | 19–30 tok/s |
| Llama 3.3 70B InstructMetaOpen weights (restricted) | Comfortable | Borderline | Q4_K_M | 11–18 tok/s |
Loads, but not recommended (7)
Technically runs. Expect a frustrating experience.
| Model | Chat | Agentic coding | Quant | Chat speed |
|---|---|---|---|---|
| gpt-oss-120bOpenAIOpen source | Borderline | Borderline | MXFP4 | 30–50 tok/s |
| Qwen3.5 122B-A10BAlibaba QwenOpen source | Borderline | Borderline | Q4_K_M | 10–17 tok/s |
| Mistral Small 4 (119B-A6.5B)Mistral AIOpen source | Borderline | Borderline | Q4_K_M | 16–25 tok/s |
| DeepSeek-V4-Flash (284B-A13B)DeepSeekOpen source | Borderline | Technically Runs | Q4_K_M | 4.2–7 tok/s |
| GLM-4.5-Air (106B-A12B)Z.aiOpen source | Borderline | Borderline | Q5_K_M | 6–11 tok/s |
| Llama 4 Scout (109B-A17B)MetaOpen weights (restricted) | Borderline | Borderline | Q4_K_M | 7–12 tok/s |
| GLM-5.3-Flash (320B-A18B)Z.aiOpen source | Borderline | Technically Runs | Q3_K_M | 3.9–7 tok/s |
Doesn't fit or isn't supported (6)
Not enough usable memory, or the runtime can't run it here.
| Model | Chat | Agentic coding | Quant | Chat speed |
|---|---|---|---|---|
| DeepSeek-V4.1-Flash (748B-A16B)DeepSeekOpen source | Does Not Fit | Does Not Fit | — | — |
| Muse Spark Open 816B-A42BMeta | Does Not Fit | Does Not Fit | — | — |
| GLM-5.3 (753B-A40B)Z.ai | Does Not Fit | Does Not Fit | — | — |
| Nemotron 3.5 Super VL 120B-A12BNVIDIA | Unsupported | Unsupported | — | — |
| Reflection Beam (501B-A23B)Reflection AIOpen source | Unsupported | Unsupported | — | — |
| Step 5 Preview 600B-A27BStepfun | Unsupported | Unsupported | — | — |
Want a specific tool, context size or workload?
The full check models your exact setup, including the apps you keep open.