Models / Qwen3.8 27B / VRAM
How much VRAM do you need for Qwen3.8 27B?
Last updated:
Weight math from the model card, plus every measured RigBase result for Qwen3.8 27B. Leave extra room for context (KV cache).
Weight size by quant
| Quant | Bits/weight | Weights only |
|---|---|---|
| Q8_0 | 8.50 | 28.7 GB |
| Q6_K | 6.56 | 22.2 GB |
| Q5_K_M | 5.70 | 19.2 GB |
| Q4_K_M | 4.89 | 16.5 GB |
| Q3_K_M | 4.00 | 13.5 GB |
| Q2_K | 3.16 | 10.7 GB |
Bits-per-weight for GGUF rows: llama.cpp quantize README. Parameter count: Ollama library tag 27b (nominal). We advise staying at Q4 or above for chat quality — see Is Q4 good enough?.
Measured speeds for Qwen3.8 27B
| Hardware | Quant | Engine | tok/s |
|---|---|---|---|
| 2x RTX 5070 Ti 16 GB · 32 GB | 3-bit GSQ | capy-flash, 2 GPUs, MTP, decode at 8k pr | 230 ⚡ |
| RTX Pro 6000 Max-Q (Blackwell, 96 GB) · 96 GB | NVFP4 (W4A16) + MTP (5 draft tokens), 256K context | vLLM 0.27.1 | 150 ⚡ |
| RTX 3090 · 24 GB | Q4_K_M (unsloth), greedy code prompt | OpenJet megakernel (MTP, 4 drafts) | 139.7 ⚡ |
| 2x RTX 5070 Ti 16 GB · 32 GB | 3-bit GSQ | capy-flash (1 GPU), MTP, decode at 8k pr | 139 ⚡ |
| RTX 5090 32GB · 32 GB | Q4_K_M, q4_0 KV | llama.cpp b10448, MTP n=2 (--spec-type d | 136.7 ⚡ |
| 2x Tesla V100 32 GB (PG500-216, sm_70) · 64 GB | Q8_0 | bnk dense engine, tensor parallel, MTP d | 104 ⚡ |
| Intel Arc Pro B70 32 GB · 32 GB | GPTQ Int4 (asym) | vLLM XPU b70 image + MTP, reasoning prom | 96 ⚡ |
| 2x RTX 3090 · 48 GB | W4A16 AutoRound + DFlash2 | vLLM TP2, 1 stream | 91.2 ⚡ |
| RTX Pro 6000 Max-Q (Blackwell, 96 GB) · 96 GB | BF16 + MTP (5 draft tokens), 256K context | vLLM 0.20.1 | 85 ⚡ |
| RTX 3090 · 24 GB | Q4_K_M (unsloth), reasoning mode | OpenJet megakernel (MTP) | 78.0 ⚡ |
| RTX 5090 · 32 GB | Q4_K_M (EPYC 9534 host) | llama.cpp 0.7.0-dev (b11669) | 75.2 |
| RTX 5090 · 32 GB | Q4_K_M | llama.cpp | 75.19 |
⚡ = speculative decoding (MTP, DFlash…). Full list: best home rig for Qwen3.8 27B.
Related
All Qwen3.8 27B builds VRAM + KV calculator By VRAM size GGUF vs AWQ vs FP8 vs NVFP4 How much system RAM?
FAQ
How much VRAM do you need for Qwen3.8 27B?
Weights alone are about 16.5 GB at Q4_K_M (27B × 4.89 bits ÷ 8). Add memory for the KV cache (context) and the runtime. The smallest-memory measured result in RigBase is 12 GB (RTX 3060) at 5.2-6.7 tok/s. Going below Q4 usually hurts quality for chat; see the Q4 guide.
What is the fastest measured speed for Qwen3.8 27B?
RTX 5090: 75.2 tok/s plain decode (Q4_K_M (EPYC 9534 host), llama.cpp 0.7.0-dev (b11669)).
Does system RAM replace VRAM?
Only for overflow. When the whole model fits in GPU or unified memory, system RAM barely matters for decode speed. See the system-RAM guide.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.