DeepSeek V4.1 Flash 1017: MIT-licensed update, vendor-run gains

DeepSeek published DeepSeek-V4.1-Flash-1017 on Hugging Face on 17 October 2026. The model card calls it a minor update to DeepSeek-V4.1-Flash, which appeared on Hugging Face on 10 September 2026, and says it improves reasoning, coding, agentic tasks and visual benchmarks while reducing task execution time and output length. The weights and repository are under the MIT licence.

What DeepSeek reports

The model card carries a benchmark table that compares the new snapshot with the previous Flash release and four rivals. The table is DeepSeek's. The card says it used the Claude Code harness for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1, and it does not say who ran the other benchmarks.

• Terminal-Bench 4.0: 42.3, up from 31.2. The table lists GPT-5.6 Sol at 39.9, GLM-5.3 at 37.9 and Opus 5 at 51.8.
• DeepSWE v1.1: 77.0, up from 74.2. The table lists Opus 5 at 74.0 and GPT-5.6 Sol at 73.0.
• ProgramBench: 29.1, up from 20.3, against 37.0 for Opus 5. NL2Repo-Bench: 68.2, up from 65.4, against 75.3 for Opus 5.
• Automation-Bench v1.0.6: 60.5, up from 54.8, against 50.3 for Opus 5. Agents' Last Exam: 35.4, up from 31.8, against 28.6 for Opus 5.
• Humanity's Last Exam without tools: 42.6, up from 36.8, against 56.3 for Opus 5.

The card's claim that the model finishes tasks faster and writes shorter outputs is stated without a figure, so a team cannot work out a cost per task from it. The card also points to a technical report, which sits in the repository of the earlier V4.1-Flash release; I have not read it.

What running it involves

The card recommends temperature 1.0 and top_p 1.0, a 1 million-token context window and max_tokens of at least 256,000. It ships no Jinja chat template. Instead, an encoding folder holds a Python reference implementation of the prompt format, and DeepSeek released deepseek-recipe, a set of Rust libraries with Python bindings, for production use. DeepSeek says that toolkit converts Messages, Chat Completions and Responses API requests into the model's prompt format and parses output back, and leaves inference, tool execution and HTTP transport to the caller.

Hugging Face's metadata for the repository counts about 763 billion parameters across FP8 and 8-bit tensors, and lists roughly 510 GB of storage. The model card does not state a parameter count, and OpenRouter's description of the model gives a different architecture summary: 552 billion in the backbone with 8 billion active on input and 16 billion on output. I could not reconcile the two, so size your hardware from the repository files, not from either summary.

Price on a reseller

OpenRouter, a third-party router, lists the snapshot at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.003 per million and a 1,048,576-token context. Its price data also carries weekday time windows at $0.30 and $1.20, double the base rate. OpenRouter says the model is served by DeepSeek under the deepseek-flash API name. That is OpenRouter's listing, not a DeepSeek price page, and it can change.

What commenters said

The Hacker News thread (item 50129045; the site is not on this publication's source list, so it is not linked) had 126 points and 15 comments when I read it. The sharpest objection was about the framing: one commenter quoted the card's phrase "minor update" next to benchmark gains and called it a hilarious sentence, and another said they were waiting for Artificial Analysis to confirm how well the benchmarks hold up. A third commenter said they wished smaller Flash models would return because running this one locally is costly. I found no independent evaluation of this snapshot at publication time.

What a team should do

• If you route traffic to V4.1 Flash today, run your own task set against the 1017 snapshot before changing the model name; the vendor table is not your workload.
• Check whether your harness matches the ones DeepSeek used for Terminal-Bench 4.0 and DeepSWE; a different harness can move these scores.
• Measure tokens per task and wall-clock time yourself. The card claims improvements on both but publishes no number.
• If you self-host, read the weight files and the inference folder first. The two parameter summaries above disagree.