← Qwen3.8 27B model page · All models
Can you run Qwen3.8 27B locally?
Qwen/Qwen3.8-27B · 27.8B parameters (dense) · license see model card · on Hugging Face since 2026-08-05
Yes. At 4-bit, Qwen3.8 27B needs about 16.8 GB for the weights (estimate from the parameter count on the model card). As of Nov 5, 2026, the cheapest US machine we track that fits is GMKtec EVO-X2 64GB / 1TB at $2,200 / €2,000, measuring 12.89 tok/s on this model.
Memory for weights (estimate)
Q4_K_M: ≈16.8 GB
Q8_0: ≈29.5 GB
From parameter count × bits per weight, same rule as our model pages. Plus room for context.
Machines that fit at 4-bit
27 of the complete machines we track
Measured results
17 public tok/s results on tracked chips
Engines and quants differ between sources.
Cheapest machines that fit
Cheapest in the US right now: GMKtec EVO-X2 64GB / 1TB at $2,200. Buy: Amazon US · Amazon EU.
Store links go through our /go/ redirects. Some are affiliate links.
Which tracked machines can run it
| Machine | Price | Measured tok/s |
|---|---|---|
| GMKtec EVO-X2 64GB / 1TB Strix Halo · 64GB | $2,200 €2,000 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Framework Desktop 64GB Strix Halo · 64GB | ≈$2,479 €2,209 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Minisforum MS-S1 MAX 64GB / 2TB Strix Halo · 64GB | $2,599 €2,799 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Geekom A9 Mega 64GB / 2TB (Ryzen AI Max+ 388) Strix Halo · 64GB | $2,599 €2,499 | no public result |
| Mac mini M5 Pro 16-core GPU Apple · 64GB | $2,699 €3,129 | no public result |
| Mac mini M5 Pro 20-core GPU Apple · 64GB | $2,899 €3,349 | 12.7 MLX 4-bit, oMLX, 1k context, 1 run |
| Framework Desktop 128GB Strix Halo · 128GB | $3,449 €3,889 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| GMKtec EVO-X2 128GB / 1TB Strix Halo · 128GB | $3,500 €3,350 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| GMKtec EVO-X2 128GB / 2TB Strix Halo · 128GB | $3,650 €3,500 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| GMKtec EVO-X3 128GB / 2TB Strix Halo · 128GB | $3,800 €3,600 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Minisforum MS-S1 MAX 128GB / 2TB Strix Halo · 128GB | $3,919 €4,319 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Geekom A9 Mega 128GB / 2TB Strix Halo · 128GB | $3,999 €3,999 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Minisforum MS-S1 MAX 128GB / 4TB Strix Halo · 128GB | $4,199 €4,599 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Beelink GTR9 Pro 128GB / 2TB Strix Halo · 128GB | $4,349 ≈€3,875 | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| Mac Studio M5 Max 40-core GPU Apple · 128GB | $5,099 €5,909 | 33.7 MLX 4-bit, oMLX, 1k context, 1 run |
| Mac Studio M5 Ultra 64-core GPU Apple · 96GB | $5,499 €6,699 | no public result |
| Dell Pro Max / Pro Precision with GB10 DGX Spark / GB10 · 128GB | $5,903 ≈€5,259 | no public result |
| ASUS Ascent GX10 (GB10) DGX Spark / GB10 · 128GB | $5,999 ≈€5,345 | no public result |
| Framework Desktop 192GB (DIY, Ryzen AI Max+ PRO 495) AMD Ryzen AI Max+ PRO 495 · 192GB | $6,799 €7,659 | no public result |
| NVIDIA DGX Spark DGX Spark / GB10 · 128GB | $6,950 ≈€6,192 | no public result |
| MacBook Pro 16" M5 Max 40-core GPU Apple · 128GB | $6,999 €7,999 | 33.7 MLX 4-bit, oMLX, 1k context, 1 run |
| Minisforum MS-S1 MAX-P495 192GB / 2TB (Ryzen AI Max+ PRO 495) AMD Ryzen AI Max+ PRO 495 · 192GB | $7,399 €7,799 | no public result |
| Mac Studio M5 Ultra 80-core GPU Apple · 256GB | $10,799 €12,529 | 49.2 MLX 4-bit, mlx-lm 0.31.3, 512-token prompt, median of 5 |
| Mac Studio M5 Ultra 80-core GPU Apple · 512GB | $22,299 €25,179 | 49.2 MLX 4-bit, mlx-lm 0.31.3, 512-token prompt, median of 5 |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395) Strix Halo · 128GB | n/a n/a | 12.89 Q4_K_M, Ollama 0.32.15 Vulkan, no MTP drafting |
| MSI EdgeXpert MS-C931 (GB10) DGX Spark / GB10 · 128GB | n/a n/a | no public result |
| Lenovo ThinkStation PGX (GB10) DGX Spark / GB10 · 128GB | n/a n/a | no public result |
tok/s is text generation from the linked public source for that chip. Fit is an estimate: weights × 1.15 + 2 GB must fit in 85% of unified memory. Full device list including GPUs: Qwen3.8 27B model page.
FAQ
Can you run Qwen3.8 27B locally?
Yes, on a machine with enough unified memory. At 4-bit the weights take about 16.8 GB (estimate), so all 27 complete machines we track that fit are listed above.
How much memory does Qwen3.8 27B need?
About 16.8 GB at Q4_K_M and 29.5 GB at Q8_0 for the weights alone, plus context. Source: parameter count on the Hugging Face model card, converted with our published memory rule.
What is the cheapest machine that can run it?
In USD, GMKtec EVO-X2 64GB / 1TB at $2,200, prices checked Nov 5, 2026.