Memory Requirements
For the recommended 4-bit GGUF, plan for 17 to 19 GB of total RAM and VRAM.
Qwen3.8-27B
Qwen3.8-27B is a hybrid thinking open model for agentic coding and chat. Run the model locally via Unsloth Dynamic GGUFs and Unsloth Desktop.
Unsloth Dynamic V3.027B parametersText + vision256K context17 to 19 GB at 4-bit
Capabilities
Text and vision input, hybrid thinking, agentic coding, nested tool calls, and up to 256K context.
Run locally
Unsloth Run
unsloth run --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
Applications
Launch a coding agent against Qwen3.8-27B with unsloth start.
Claude Code
unsloth start claude --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
OpenAI Codex
unsloth start codex --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
Hermes Agent
unsloth start hermes --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
OpenClaw
unsloth start openclaw --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
OpenCode
unsloth start opencode --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
Run options
Use the GGUF build on most computers or NVFP4 on NVIDIA Blackwell.
GGUF 4-bit
RecommendedBest balance of local quality and memory use.
Unsloth Run
unsloth run --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XLUse Unsloth Desktop to search, download, and tune models automatically.
NVFP4
About 2.5x fasterOptimized for RTX 50 series, DGX Spark, B200, and B300.
vLLM
vllm serve unsloth/Qwen3.8-27B-NVFP4SGLang
python -m sglang.launch_server --model-path unsloth/Qwen3.8-27B-NVFP4 --speculative-algorithm NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4FP8 KV cache calibration supports longer context.
Memory and settings
Unsloth applies the recommended inference values automatically.
Memory requirements
| Quant | 2-bit | 3-bit | 4-bit | 6-bit | 8-bit | BF16 |
|---|---|---|---|---|---|---|
| Total memory | 11 to 13 GB | 13 to 16 GB | 17 to 19 GB | 24 GB | 31 GB | 56 GB |
Use RAM plus VRAM close to the quant size for best speed. Less memory can work with slower disk offloading.
Run, train, and connect locally with Unsloth Desktop
Search and download GGUF or safetensor models, use self-healing tool calls, run web search and code, and let Unsloth tune inference automatically on macOS, Windows, or Linux.
