DeepSeek has published DeepSeek-V4.1-Flash-1017 on Hugging Face. It calls it a minor update to V4.1-Flash. The model card says it improves reasoning, coding, agentic tasks and visual benchmarks. It also says the model finishes tasks faster and writes shorter outputs, so you pay less per task.
The model card's benchmark table compares it with the previous Flash, and with K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Terminal-Bench 4.0 goes from 31.2 to 42.3. DeepSWE v1.1 goes from 74.2 to 77.0. ProgramBench goes from 20.3 to 29.1. DeepSeek's own numbers put it ahead of every model in the table except Opus 5 on several tests, and ahead of all of them on DeepSWE v1.1, Automation-Bench and Agents' Last Exam. Opus 5 is still clearly ahead on Terminal-Bench, ProgramBench and HLE.
The weights are under the MIT license, and the card lists 763B parameters. Recommended settings are temperature 1.0, top_p 1.0, a 1M-token context window and max_tokens of 256K or more. There is no Jinja chat template. DeepSeek ships a Python reference encoder and a Rust toolkit called deepseek-recipe, which handles prompt formatting, tool calls, thinking and images. At the time of the fetch, no Inference Provider on Hugging Face hosted it. Quantized versions were already listed for llama.cpp, Ollama and LM Studio.
My take
The MIT license matters most to me. You can use, change and ship this commercially without asking anyone. For a small team, that is a very different position from renting a closed model.
I'd treat the benchmark table with some care. Every number comes from DeepSeek itself, on its own choice of harnesses. The card says it used the Claude Code harness for Terminal-Bench and Mini-SWE-V2 for DeepSWE. The jump over its own previous version is large for a "minor" update, so I'd like independent runs before believing it. Beating Opus 5 on a few tests doesn't make it better overall. Opus 5 leads several of the others by a wide margin.
The practical catch is size. 763B parameters is not something you run on a laptop or a single cheap GPU. The quantizations will help, but most of us will use it through an API, and the card says no provider hosts it yet. If you build from Nigeria, as I do, hosted access and pricing will decide whether this is useful, not the license. I'd wait for a provider or two to list it and compare the real price per finished task. The "shorter outputs" claim could make that price lower, but only a real bill will show it.
If you build agents, the encoding folder and deepseek-recipe are the first things I'd read. Prompt formatting and tool-call parsing are where open models tend to break in subtle ways. A maintained toolkit that covers them is a real help, and the missing chat template is the mistake that could catch you out.
Sources
This post was written with AI assistance from the sources above.
Comments
What do you think? Drop your thoughts below.
Loading comments…