#2 Qwen3.8-27B
Strongest model that runs on one 24 GB GPU, Apache-2.0, huge context. Default pick for most local RAG builders.
- Version
- Qwen3.8-27B
- Maker
- Alibaba Qwen
- Release date
- 2026-08-05 (Hugging Face repo creation date unless noted)
- Architecture
- Dense, image+video input
- Total params
- 27B (27.8B per HF)
- Active params
- 27B
- Context length
- 262,144 native, extensible to ~1M
- License
- Apache-2.0
- Commercial notes
- Fully commercial
- Variants
- Qwen3.8-27B
- Qwen3.8-27B-FP8
- unsloth GGUF
- Qwen3.8-Flash-Next 125B/6B-active (qwen-community-1.0 license)
- Memory
- BF16/native: ~54 GB (Yotta Labs)
FP8: no published figure yet
Q8: no published figure yet
Q4: ~16 GB (Yotta Labs; 24 GB GPU class) - Hardware
- Single Spark (128GB): Yes, BF16 fits
Dual Spark (256GB): Yes
Quad Spark (~512GB): Yes
24GB GPU: Yes
128GB Mac / Strix: Yes - Tokens/sec
- apple_m5_ultra_bf16: 19.69 tok/s (Apple, 8K ctx, via LLMCheck)
- apple_m3_ultra_bf16: 12.46 tok/s (Apple, via LLMCheck)
- apple_m5_max_q4: ~29 tok/s ESTIMATE (LLMCheck)
- Document Q&A strengths
- 262K native context
- Thinking on/off per request
- Best dense single-GPU model; tied DeepSeek V4 Flash on old AA scale
- Embedding / reranker pairing
- no published figure yet
- Runtimes
- Ollama: listed (qwen3.8)
llama.cpp: Yes (unsloth GGUF)
vLLM: Yes (model card)
LM Studio: no published figure yet unverified - Hugging Face
- https://huggingface.co/Qwen/Qwen3.8-27B
DGX Spark configs it fits
- Single Spark (128GB): Yes, BF16 fits
- Dual Spark (256GB): Yes
- Quad Spark (~512GB): Yes
Spark fit is estimated from published memory sizes, not tested on Spark hardware.
Sources
- https://huggingface.co/Qwen/Qwen3.8-27B
- https://www.yottalabs.ai/post/best-chinese-llm-models-2026-deepseek-qwen-glm-kimi-compared
- https://llmcheck.net/blog/state-of-open-source-local-llms-october-2026/
- https://ollama.com/library/qwen3.8
Last updated Oct 24, 2026. Fields tagged unverified lack a primary source. "No published figure yet" means we found no source, and we don't guess. Spark fit is estimated from published memory sizes, not tested on Spark hardware. Ollama listings may be cloud-only tags; check this for Kimi K3 and DeepSeek V4.1 before pulling.
The RAGShelf weekly
New local RAG tools, releases, and tested setups. One email a week, no spam.