DeepSeek-V4.1-Flash-1017 is presented on Hugging Face as a minor update to DeepSeek-V4.1-Flash. The image-text-to-text model is tagged for Transformers, Safetensors, 8-bit precision, and fp8. It has 763B params and supports BF16, F32, F8_E4M3, and I8 tensor types. The repository uses the MIT License.

terminal
pip install vllm
vllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"

DeepSeek reports gains in reasoning, coding, agentic work, and visual benchmarks. The update also reduces task execution time and output length, which the model card positions as a better performance and cost tradeoff across agent harnesses.

The benchmark comparison uses DS-V4.1-Flash, K3, GLM-5.3, GPT-5.6 Sol, and Opus 5 as reference columns. On HLE w/o tools, DeepSeek-V4.1-Flash-1017 scores 42.6, or 44.0 on the marked subset, compared with 36.8 and 39.1 for DS-V4.1-Flash, 43.5 for K3, 42.0 for GLM-5.3, 44.5 for GPT-5.6 Sol, and 56.3 for Opus 5. On Terminal-Bench 4.0, the scores are 42.3, 31.2, 12.6, 37.9, 39.9, and 51.8 in that order.

DeepSWE v1.1 records 77.0 for the new model, followed by 74.2, 67.5, 66.9, 73.0, and 74.0. ProgramBench records 29.1, 20.3, 17.5, 19.0, 23.0, and 37.0. NL2Repo-Bench records 68.2, 65.4, 58.0, 58.0, 56.8, and 75.3.

On Automation-Bench v1.0.6, the figures are 60.5, 54.8, 46.7, 48.8, 45.8, and 50.3. Agents’ Last Exam records 35.4, 31.8, 27.6, 28.5, 26.7, and 28.6. Apex Bench v1.1 records 72.1, 64.6, 47.8, 56.6, 54.3, and 65.8.

For tool-assisted tests, Chartography records 83.0, 78.9, 68.1, no value for GLM-5.3, 79.9, and 84.0. BabyVision records 92.5, 89.6, 85.7, no GLM-5.3 value, 88.9, and 94.1. ZeroBench-main records 51.0, 49.0, 41.0, no GLM-5.3 value, 53.0, and 52.0. The dagger on HLE identifies the text-only subset. DeepSeek used the Claude Code harness for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1.

The release has no Jinja-format chat template. Its encoding folder contains the self-contained Python reference implementation encoding.py with tests for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, system messages introduced during a conversation, and interleaved image content.

DeepSeek also provides deepseek-recipe, a set of Rust libraries with Python bindings. The toolkit converts Messages, Chat Completions, and Responses API requests into the Conversation format. It can encode DeepSeek V4 and V4.1 prompts or token IDs and parse complete or streamed responses, including thinking, tool calls, images, and generation settings. Model inference, tool execution, and HTTP transport remain the caller’s responsibility.

The model card points to the inference folder for weight conversion and local execution. It documents Transformers, vLLM, SGLang, Docker Model Runner, Google Colab, Kaggle, and local apps. Quantizations are available for llama.cpp, Ollama, LM Studio, and compatible applications. Recommended settings are temperature 1.0, top_p 1.0, a 1M-token context window, and max_tokens of at least 256K.

The supplied citation names DeepSeek-AI as author, uses the title DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression, and gives the year 2026.