Xolotl (Strata based) - DeepSeek-V4.1-Flash (552B) on a single 4090 at ~27 tok/s, Uncensored 1017 quant coming in a couple of days
I've been building an inference engine based off Strata called Xolotl that runs DeepSeek-V4.1-Flash, all 552B of it, on one gaming PC.
Hugging FaceHumanEvaldeepseek-ai/DeepSeek-V4.1-FlashQwen/Qwen3.8-Flash-NextQwen/Qwen3.8-27Bzai-org/GLM-5.3-Flash
Proof1 community thread
Key points
- Hey all, been meaning to post this for a while.
- My current quant (uncensored V4.1, 152 GiB) is here:
- The 1017 update: I'm downloading the new DeepSeek-V4.1-Flash-1017 right now.
- It's the same architecture and config as V4.1 (I checked, identical config and tokenizer), So it should run at the same throughput as the current one.
Sources (1)
- [1]Xolotl (Strata based) - DeepSeek-V4.1-Flash (552B) on a single 4090 at ~27 tok/s, Uncensored 1017 quant coming in a couple of daysr/LocalLLaMA (top, daily) · Oct 17, 09:40 PM
I've been building an inference engine based off Strata called Xolotl that runs DeepSeek-V4.1-Flash, all 552B of it, on one gaming PC.
Hey all, been meaning to post this for a while.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026deepseek-ai/DeepSeek-V4.1-Flash-1017
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026capy-flash: Qwen3.8-Flash-Next (125B MoE) in full NVFP4 on two 16 GB consumer GPUs
- Oct 15, 2026QUASAR 4-bit QAT checkpoints for Qwen3.5, Gemma 4 and Muse Glimmer: closer to BF16 than Google’s and NVIDIA’s 4-bit releases
- Oct 15, 2026Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
