---

license: apache-2.0

base_model: Qwen/Qwen3.8-27B

base_model_relation: quantized

tags:

- gguf

- ollama

- llama.cpp

- apple-silicon

- qwen3.8

- q8_0

---

Qwen3.8-27B — GGUF Q8_0

Single-file Q8_0 (8-bit, near-lossless) GGUF of

Qwen/Qwen3.8-27B — 28.9 GB.

Quantized by AtomicChat;

re-hosted here alongside our Apple-Silicon serving research for one-command use.

Run it

# Ollama (from the Ollama registry — easiest)
ollama run bluehawana/qwen3.8-27b-q8

# Ollama (straight from this repo)
ollama run hf.co/bluehawana/Qwen3.8-27B-Q8-GGUF

# llama.cpp / LM Studio / Jan: download Qwen3.8-27B-Q8_0.gguf directly

Needs ≥48 GB unified memory on Apple Silicon (comfortable at 64 GB+).

Concurrent serving on a Mac

This model serves 16 concurrent requests with zero errors on an M-series

128 GB Mac — benchmarks across Ollama / oMLX / SGLang-MLX, plus the SGLang

patches that make it possible (upstream PR

sgl-project/sglang#35137):

→ bluehawana/qwen3.8-27b-apple-silicon-concurrency

Quick concurrent Ollama serving:

OLLAMA_HOST=127.0.0.1:11500 OLLAMA_NUM_PARALLEL=16 OLLAMA_CONTEXT_LENGTH=8192 ollama serve
# OpenAI-compatible endpoint: http://127.0.0.1:11500/v1

Credits: base model © Qwen (Apache-2.0) · Q8_0 quant by AtomicChat · benchmarks & patches by bluehawana.