#3 DeepSeek V4.1 Flash
Top benchmark scores and MIT, but no mainstream local runtime yet and too big for quad Spark natively; V4-Flash-0731 is the practical local DeepSeek.
- Version
- DeepSeek-V4.1-Flash-1017
- Maker
- DeepSeek
- Release date
- 2026-10-17 (V4.1 Flash 2026-09-10) (Hugging Face repo creation date unless noted)
- Architecture
- Causal encoder-decoder MoE with Engram memory, native image input
- Total params
- 552B (Yotta Labs; HF safetensors total ~763B incl. extras)
- Active params
- 8B input / 16B output
- Context length
- 1M
- License
- MIT
- Commercial notes
- Fully commercial
- Variants
- V4.1-Flash-1017
- V4.1-Flash
- V4-Flash-0731 (284B/~13B, 166.9 GB, easier to host)
- V4-Pro-0813 (1.6T/49B)
- Memory
- BF16/native: 510 GB on disk, ~614 GB GPU memory (Yotta Labs)
FP8: no published figure yet
Q8: no published figure yet
Q4: no published figure yet
V4-Flash (older) runs at 2-bit on a 128 GB Mac per LLMCheck - Hardware
- Single Spark (128GB): No (use V4-Flash-0731 at 2-bit)
Dual Spark (256GB): No
Quad Spark (~512GB): Only if quantized, untested unverified
24GB GPU: No
128GB Mac / Strix: No (V4-Flash 2-bit yes)
Native weights exceed 512 GB - Tokens/sec
- m3_ultra_512gb_omlx: 20.23 tok/s community (LLMCheck)
- v4_flash_m5max_2bit: 39 tok/s community (LLMCheck)
- api_artificial_analysis: 186 tok/s (hosted, not local)
- Document Q&A strengths
- 1M context, tiny 890-byte/token KV cache
- #1 open model on Artificial Analysis Intelligence Index (45.1)
- MIT
- Embedding / reranker pairing
- no published figure yet
- Runtimes
- Ollama: listed (deepseek-v4.1-flash; may be cloud tag)
llama.cpp: No upstream support yet (open PR)
vLLM: no published figure yet unverified
LM Studio: no published figure yet - Hugging Face
- https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
⚠️ The Ollama listing for this model may be a cloud-only tag, not local weights. Check before pulling.
DGX Spark configs it fits
- Quad Spark (~512GB): Only if quantized, untested unverified
Spark fit is estimated from published memory sizes, not tested on Spark hardware.
Sources
- https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
- 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://theopenweights.com/leaderboards/llm-intelligence
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.