Can You Run Qwen3.8-27B? VRAM Requirements
Qwen3.8-27B has 27B total parameters. Running it at Q4_K_M (GGUF) with 16K context takes roughly 17.3 GB of memory (15.3 GB weights + 0.5 GB KV cache + runtime overhead). The minimum viable hardware is about RTX 3090 / 4090 24GB. All numbers are planning estimates, verified 2026-09-05.
VRAM by quantization (16K context)
| Quantization | Weights | Total needed | Minimum hardware |
|---|---|---|---|
| FP16 / BF16 | 50.3 GB | 54.8 GB | Mac 64GB unified |
| FP8 | 25.1 GB | 27.7 GB | RTX 5090 32GB |
| Q8_0 (GGUF) | 26.7 GB | 29.3 GB | RTX 5090 32GB |
| Q6_K (GGUF) | 20.6 GB | 22.8 GB | RTX 5090 32GB |
| Q5_K_M (GGUF) | 17.9 GB | 19.9 GB | RTX 3090 / 4090 24GB |
| MXFP4 | 13.3 GB | 15.3 GB | RTX 3090 / 4090 24GB |
| NVFP4 | 13.3 GB | 15.3 GB | RTX 3090 / 4090 24GB |
| Q4_K_M (GGUF)recommended | 15.3 GB | 17.3 GB | RTX 3090 / 4090 24GB |
| Q3_K_M (GGUF) | 11.8 GB | 13.8 GB | RTX 4060 Ti 16GB |
| Q2_K (GGUF) | 8.8 GB | 10.8 GB | RTX 3060 12GB |
How context length changes memory (Q4_K_M (GGUF))
| Context | KV cache | Total needed |
|---|---|---|
| 4K | 0.1 GB | 17.0 GB |
| 16K | 0.5 GB | 17.3 GB |
| 64K | 2.0 GB | 18.8 GB |
| 128K | 4.0 GB | 20.8 GB |
| 256K | 8.0 GB | 24.8 GB |
KV cache grows linearly with context; FP8 KV cache halves it again (try it in the calculator below).
Measured speeds for Qwen3.8-27B
Measured by us; every row links to the source article. The tables above are estimates — these are real readings.
| Machine | Quant / runtime | Setup | Decode tok/s | Prefill | Peak mem | Source |
|---|---|---|---|---|---|---|
| Mac mini M4 24GB | 4-bit (MLX) MLX | Plain autoregressive baseline, 3 runs | 6.4–6.5 | — | — | article 2026-08-29 |
| Mac mini M4 24GB | 4-bit (MLX) MLX | DFlash 2 draft model (2B, 4-bit) speculative decoding, block-size 3 block-size 8 drops to 1.11x; an 8-bit draft is slower (1.62x) | 11.6–12.21.8–1.9x | — | 19.4GB | article 2026-08-29 |
| Mac mini M4 24GB | UD-Q4_K_S (GGUF, 15.4GB) llama.cpp b96806d (Metal) | Plain autoregressive baseline (prose / code) | 5.9–6 | — | 15.4GB | article 2026-09-02 |
| Mac mini M4 24GB | UD-Q4_K_S (GGUF) llama.cpp b96806d (Metal) | Native MTP head speculative decoding, n-max 2 Prose -24%, code -4%: a net slowdown on Metal | 4.5–5.70.76–0.96x | — | 16GB | article 2026-09-02 |
| Mac mini M4 24GB | UD-Q4_K_S (GGUF) llama.cpp b96806d (Metal) | DFlash 2 GGUF draft (1.14GB), n-max 3 The recommended n-max 7 hits a reproducible Metal OOM at 8K context | 3–4.60.5–0.77x | — | — | article 2026-09-03 |
| Mac mini M4 24GB | UD-Q4_K_S (GGUF) llama.cpp llama-server b11376 | Defaults (--fit caps context at 30K, 4 slots) | 6.3 | 56 | — | article 2026-10-03 |
| Mac mini M4 24GB | UD-Q4_K_S (GGUF, 自导入) Ollama 0.35.1 | Defaults (4096 context) Ollama 0.35's inference process is its bundled llama-server; same flags, same speed | 6.3 | 57 | — | article 2026-10-03 |
Browse by machine: Mac mini M4 24GB
Customize the estimate
LLM VRAM Calculator — Can I Run It?
Estimate how much GPU memory (VRAM) you need to run open-weight LLMs like Kimi K3, DeepSeek R1, Qwen3, or Llama locally. Pick a model, quantization, and context length — the calculator adds up weights, KV cache, and runtime overhead, then shows which GPUs or Macs can fit it. Runs entirely in your browser: no upload, no signup.
27B total params · max 256K context
Recommended balance · ~0.61 bytes/param
Estimated memory needed
20.8 GB
Will it fit?
Assumes ~92% of device memory is usable. Multi-GPU counts are for tensor/pipeline parallel serving (vLLM, SGLang); Apple Silicon uses unified memory via llama.cpp or MLX.
FAQ
How much VRAM does Qwen3.8-27B need?
About 17.3 GB at the recommended Q4_K_M (GGUF) quantization with 16K context: 15.3 GB for weights, 0.5 GB for KV cache, plus runtime overhead. Longer context grows the KV cache.
What is the minimum hardware for Qwen3.8-27B?
Roughly RTX 3090 / 4090 24GB, assuming 92% of device memory is usable. Lower-end setups can try more aggressive quantization (Q3/Q2) at a noticeable quality cost.
Can I run Qwen3.8-27B on a Mac?
Yes. At Q4_K_M (GGUF) with 16K context it needs about 17.3 GB, so a Mac 24GB unified or larger fits it via llama.cpp or MLX. Note macOS only lets the GPU wire roughly 2/3-3/4 of RAM by default; near the limit, raise iogpu.wired_limit_mb or use a smaller quant.
Which quantization should I use for Qwen3.8-27B?
Q4_K_M is the recommended balance of size and quality; with headroom, Q6_K or Q8_0 reduce quality loss further.