DeepSeek has released DeepSeek-V4.1-Flash-1017, a 763-billion-parameter model whose weights and code carry the MIT licence, and its own benchmark table shows a model that has caught the frontier at doing work but not at thinking hard.

Start with the licence, because it is the fact most likely to change a decision. MIT carries no revenue threshold and no acceptable-use rider, the two conditions that make many “open” model releases awkward for a company with real customers. You can download it, modify it, and sell products built on it without asking. The model card, published on Hugging Face, shows 17,924 downloads in the past month, plus 65 quantized versions, eight fine-tunes and one merge from outside developers.

The card opens by calling the release a “minor update” to the earlier DeepSeek-V4.1-Flash, then claims “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” Both statements hold, because the comparison is with the predecessor. Terminal-Bench 4.0, a test of working in a command line, rose from 31.2 to 42.3. ProgramBench went from 20.3 to 29.1. The hardest reasoning exam, Humanity’s Last Exam without tools, climbed from 36.8 to 42.6.

Set the same table against the frontier and the picture splits. Opus 5 still scores 56.3 on that reasoning exam. It also leads on Terminal-Bench, 51.8 to 42.3, and on ProgramBench, 37.0 to 29.1. It wins NL2Repo-Bench, which asks a model to rebuild a code repository, 75.3 to 68.2, and two image-understanding tests, Chartography (84.0 to 83.0) and BabyVision (94.1 to 92.5). On ZeroBench, another image test, Opus 5 scores 52.0 to DeepSeek’s 51.0, and OpenAI’s GPT-5.6 Sol scores 53.0. GPT-5.6 Sol also edges DeepSeek on the reasoning exam, 44.5 to 42.6.

The wins sit somewhere else. DeepSeek leads every model in the table on four tests, and all four measure whether a model can finish multi-step tasks on its own. DeepSWE v1.1 puts it at 77.0 against 74.0 for Opus 5. Automation-Bench has it at 60.5 against 50.3, a ten-point gap. Agents’ Last Exam reads 35.4 against 28.6, and Apex Bench 72.1 against 65.8. Finishing tasks is what teams actually pay for, so this is the result worth taking seriously.

Treat the numbers as DeepSeek’s. The card names no independent evaluator, and the company chose the software wrapped around the model for two of the tests: Claude Code for Terminal-Bench and Mini-SWE-V2 for DeepSWE. The card also says the model cuts task time and output length, which would lower cost, but it gives no figures for either.

The practical details are specific. The context window is one million tokens, and DeepSeek recommends an output budget of at least 256,000 tokens. The page documents serving through vLLM, SGLang, Transformers and Docker, with quantized builds for llama.cpp, Ollama and LM Studio. One cost for builders: the release ships no standard Jinja chat template, the file most serving tools use to format a conversation. DeepSeek supplies a Python reference encoder and a Rust library called deepseek-recipe instead, so existing pipelines need an adapter.

One thing the card leaves out is the memory-use figures now circulating, including 890 bytes per token and a footprint about one eighth of its predecessor’s. None of that appears on the page. The only trace is the title of a technical report it cites, “Pushing the Limits of KV Cache Compression,” which this piece did not read. Treat those numbers as claims from that report until someone checks them.

For a team choosing a base model this quarter, the test is narrow and cheap to run: put it on your own multi-step workflow, with your own tooling, next to the closed model you pay for today. If it matches there, the licence leaves one question, which is what it costs to host.

Based on DeepSeek’s model card for DeepSeek-V4.1-Flash-1017 on Hugging Face, read 19 October 2026 (the page shows no publication date).