---
title: DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2
canonical_url: "https://www.modelscope.cn/models/manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2"
md_url: "https://www.modelscope.cn/models/manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2.md"
repository: manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2
last_updated: 2026-11-02
license: mit
pipeline_tag: text-generation
tasks:
  - text-generation
base_model:
  - deepseek-ai/DeepSeek-V4.1-Flash-1017
base_model_relation: finetune
parameters: 308.4B
tensor_type:
  - BF16
  - F8_E4M3
  - F8_E8M0
  - F32
  - F16
  - I32
  - I16
library_name:
  - safetensors
downloads: 52
stars: 0
---

# DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2

> DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2 - manateelazycat 在 ModelScope 开源的模型。DeepInfer native scheduler and Thor EXL3 TP2 runtime for DeepSeek V4.1 Flash 1017; independent LPK artifacts.

manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2 是 ModelScope 魔搭社区上的 308.4B 参数text-generation模型，采用 mit 许可，基于 deepseek-ai/DeepSeek-V4.1-Flash-1017 构建。

- **Repository**: manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2
- **License**: mit
- **Tasks**: text-generation
- **Parameters**: 308.4B
- **Base model**: deepseek-ai/DeepSeek-V4.1-Flash-1017
- **Downloads**: 52
- **Stars**: 0
- **Last updated**: 2026-11-02

Source: https://www.modelscope.cn/models/manateelazycat/DeepSeek-V4.1-Flash-1017-DeepInfer-Thor-TP2

---

# DeepSeek V4.1 Flash 1017 — DeepInfer Thor TP2

Independent AI Pod package: `cloud.lazycat.aipod.deepseek-v41-1017-deepinfer-tp2`.

This repository packages the DeepInfer native C++ scheduler with the custom Thor EXL3 TP2 backend. TensorFold CUDA/model components are used; TensorFold's serving engine is not instantiated.

The EXL3 weights are copied byte-for-byte from the accepted Eastham 1017 reference environment. Engram objects are byte-identical to DeepSeek's original 1017 checkpoint. The inference image contains no weights.

The qualified reference profile uses two NVIDIA Thor T5000 machines, an 8192-token context and eight request slots. The current API supports text and deterministic generation (temperature=0); visual input, tool calling and structured output are not advertised.

Publication and installation qualification are in progress. The reference benchmark is not an installed-LPK performance result. Use only the fixed revisions and hashes pinned by a completed release; the branch may contain candidates.

The final LPK, corresponding source and patches, dependency licenses and acceptance evidence will accompany the completed release on both model hubs.