DeepSeek has released the V4.1-Flash-1017 weights on Hugging Face (MIT license, image and text input). The repository includes a Python reference implementation for prompt encoding and a minimal reference inference code, but does not include a Jinja-formatted chat template.
Downloads 116770
Confidence: High (80) · Event time 2026-10-16
Reasoning
- 10-17 release (recent development, not within the past 24 hours); 10-27 has reported the model release, but this entry focuses specifically on the weights repository itself, as local deployment users are most prone to errors in prompt formatting.
- Officially, the prompt format for V4.1 differs from V4 in three ways: an extra space is added before the DSML tag name for tool calls; reasoning effort is changed to a numerical value between 1 and 100 (low=50, high=75 is default, max=100); and system messages can be inserted mid-conversation.
- For production use, the officially released deepseek-recipe (Rust library + Python bindings) can be used to convert Messages, Chat Completions, and Responses requests into V4 / V4.1 prompts or tokens, and to parse thinking, tool calls, and image outputs.
- The official model card lists improvements over the previous V4.1-Flash version, such as Terminal-Bench 4.0 increasing from 31.2 to 42.3 and ProgramBench increasing from 20.3 to 29.1 (officially stated; Terminal-Bench was tested using the Claude Code framework).
Usage
- Recommended sampling: temperature 1.0, top_p 1.0, context 1M tokens, max_tokens no less than 256K (official recommendation).
- Prompt construction: Use encode_messages(messages, thinking_mode="thinking", reasoning_effort=75) from encoding/encoding.py; parse model output using parse_message_from_completion_text; use the <｜deepseek_image｜> placeholder for images.
- Reference inference: In the inference directory, run pip install -r requirements.txt, use convert.py to convert weights according to the tensor parallelism count (example MP=8, --expert-dtype fp4), then run torchrun --nproc-per-node 8 generate.py --interactive for conversation.
Risks
- The official documentation for the inference directory explicitly states it is a "readable reference implementation, not a production serving engine," and the examples are written for 8-way tensor parallelism; to run locally on a single GPU, look for community quantized versions or engines like Delta.
- There is no Jinja template; inference frameworks that directly apply generic chat templates may incorrectly construct DSML tags or reasoning effort prefixes; all benchmark comparison figures are self-tested by the official team.
Open · Model repository encoding/README.md · Model repository inference/README.md · DeepSeek API Documentation · News 2026/10/17