Run Qwen3.8 27B locally: memory, speed and cost
Qwen3.8 27B has 27 billion parameters (dense). At Q4_K_M it needs about 20 GB with 32K tokens of context, so the cheapest hardware here that runs it is the Intel Arc Pro B70 (about 21 tokens a second). It scores 152 on Epoch AI’s Capabilities Index. Through an API it costs $0.039 / $1.25 per 1M tokens.
Memory needed
Weights plus context cache
| Quantisation | 8K context | 32K context | 128K context |
|---|---|---|---|
| Q4_K_Musual choice | 18 GB | 20 GB | 26 GB |
| Q5_K_M | 21 GB | 23 GB | 29 GB |
| Q8_0 | 31 GB | 32 GB | 39 GB |
| BF16 | 57 GB | 58 GB | 65 GB |
Estimates: weights at the stated bits per weight, plus the context (KV) cache, plus about 1 GB of overhead. Lower-bit quantisations save memory at some cost in quality.
Qwen3.8 27B on each device
Q4_K_M, 32K context
| Hardware | Memory | Fits? | Speed (estimate) |
|---|---|---|---|
| RTX 5090 | 32 GB | Fits | 75 tok/s |
| RTX 5080 | 16 GB | Partly, with system RAM | 10.1 tok/s |
| RTX 5070 Ti | 16 GB | Partly, with system RAM | 10.0 tok/s |
| RTX 5060 Ti 16 GB | 16 GB | Partly, with system RAM | 8.5 tok/s |
| RTX 4090 (used) | 24 GB | Fits | 44 tok/s |
| RTX 3090 (used) | 24 GB | Fits | 41 tok/s |
| RTX 3060 12 GB | 12 GB | Partly, with system RAM | 5.5 tok/s |
| RTX PRO 6000 Blackwell | 96 GB | Fits | 75 tok/s |
| Radeon RX 9070 XT | 16 GB | Partly, with system RAM | 9.3 tok/s |
| Radeon RX 7900 XTX | 24 GB | Fits | 42 tok/s |
| Radeon AI PRO R9700 | 32 GB | Fits | 30 tok/s |
| Intel Arc Pro B70 | 32 GB | Fits | 21 tok/s |
| Mac Studio M5 Max 128 GB | 128 GB | Fits | 27 tok/s |
| Mac Studio M5 Ultra 512 GB | 512 GB | Fits | 34 tok/s |
| Ryzen AI Max+ 395 128 GB | 128 GB | Fits | 10.9 tok/s |
| NVIDIA DGX Spark | 128 GB | Fits | 11.6 tok/s |
Generation speed estimated from memory bandwidth, calibrated against published llama.cpp results. Fastest here: RTX 5090, about 75 tokens a second. Try other settings in the calculator.
How capable is it?
Capability 152 (range 150–154) on the Epoch Capabilities Index. API models with a similar score: GLM-5.3 Flash (153, $0.15 / $0.50) and Mistral Large 4 (151, $0.70 / $2.10). Similar local models: Qwen3.5 397B-A17B (147), Qwen3.6 35B-A3B (144), Gemma 4 31B (143).
Source: Epoch Capabilities Index by Epoch AI, CC BY 4.0, read 4 Nov 2026.
Or use it through an API
$0.039 / $1.25 per 1M input / output tokens through OpenRouter (Qwen3.8 27B hosted prices). A customer chat assistant of 100M input and 20M output tokens a month costs about $28.90, so local hardware pays for itself only with heavy use.
Questions
How much VRAM does Qwen3.8 27B need?
About 18 GB at Q4_K_M with 8K tokens of context, 20 GB with 32K and 26 GB with 128K. At 8-bit (Q8_0) it needs about 31 GB with 8K context.
Can Qwen3.8 27B run on a Mac?
Yes. A Mac Studio M5 Max with 128GB runs it with 32K context at an estimated 27 tokens a second.
How fast is Qwen3.8 27B on an RTX 4090?
About 44 tokens a second with 32K context, an estimate from the card’s memory bandwidth.
How capable is Qwen3.8 27B?
It scores 152 on Epoch AI’s Capabilities Index (range 150–154), close to GLM-5.3 Flash (153) and Mistral Large 4 (151) among API models.
What does Qwen3.8 27B cost through an API?
$0.039 / $1.25 per 1M input / output tokens through OpenRouter, the cheapest host we track. A customer chat assistant of 100M input and 20M output tokens a month would cost about $28.90.