DeepSeek V4.1 Flash 1017 (Max) Lands at No. 2 on Intelligence Index

DeepSeek V4.1 Flash 1017 (Max) sits second on Artificial Analysis's intelligence ranking, according to the model page that surfaced on the Hacker News front page on 19 October 2026. The listing matters because it describes an MIT-licensed, open-weights model with 552 billion total parameters that scores near the top of a widely cited index while listing at $0.30 per million input tokens — a combination that shifts the arithmetic for anyone choosing a model this quarter. ## What does the Artificial Analysis page actually report? The page, published by Artificial Analysis at [artificialanalysis.ai/models/deepseek-v4-1-flash-1017](https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017), labels DeepSeek V4.1 Flash 1017 (Max) an open-weights model released in October 2026. It ranks #2 of 121 on the Artificial Analysis Intelligence Index with a score of 45, against a median of 19 for comparable models, and the page awards it four out of four units for intelligence. It also places #6 of 79 on speed at 230.4 output tokens per second, #21 of 43 on cost, and #39 of 47 on verbosity — that last ranking is the one to keep in mind. ## What is actually inside the model? The technical specifications list 552B total parameters with 16B active parameters, which the page defines as the number of parameters active per token during inference. That gap is the architecture's central claim: a large pool of weights, a small slice engaged for any given token. The context window is 1M tokens, which Artificial Analysis equates to roughly 1,500 A4 pages of 12-point Arial text. Input modality is text and image; output is text only. The license is MIT and the model weights are hosted on Hugging Face. The page notes that it displays the reasoning version of the model and that a non-reasoning variant may also exist. ## How is it priced, and what does one task cost? According to the listing, pricing is $0.30 per 1M input tokens and $1.20 per 1M output tokens, with a 98% cache discount. Artificial Analysis describes both figures as moderately priced, against medians of $0.43 for input and $1.68 for output. Weighted across the evaluations, the model costs $0.26 per Intelligence Index task. That per-task figure is the more useful number for budgeting, because it folds in how much the model writes as well as what it charges per token.

Why does verbosity get its own section?

Because it is the counterweight to the headline price. On the Intelligence Index, the model generated 210M output tokens, against a median of 140M — the page calls this very verbose. Output tokens cost four times input tokens here, so a model that writes substantially more than its peers gives back part of its per-token advantage. For anyone running agents that loop, retry and narrate their reasoning, token volume is often the difference between a cheap experiment and an expensive one. The $0.26 per-task figure already reflects this, which is why it is worth quoting alongside the raw rates rather than instead of them. ## How should the benchmark caveats be read? The Intelligence Index version cited is v4.3.2, which the page says incorporates ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. Two of those, SciCode and CritPt, are marked "under review" on the page, which is a reason to treat the composite score as provisional rather than final. Comparisons are also class-based: open-weights models are compared only with other open-weights models of the same size class, where anything above 150B parameters counts as Large, and reasoning models are compared across both reasoning and non-reasoning models. A #2 rank is therefore a rank inside a defined pool of 121 models, not a verdict over every model in existence. ## What does this mean in practice? Three practical readings. First, the top of the open-weights tier is now competitive on a published composite index, not just on vendor-selected benchmarks — and MIT licensing plus weights on Hugging Face removes most legal friction for teams that want to self-host or fine-tune. Second, the 16B active parameter figure is what makes serving economics interesting; it is not a claim that the model is small, since 552B total parameters is not a laptop workload. Third, a 1M-token context with image input makes long-document and screenshot-heavy work plausible in a single pass, though the verbosity finding suggests you should measure output length, not just answer quality, before committing a pipeline to it. ## Where this fits in Grafy Leaderboards are a starting point, not an answer, and Grafy is built around the part that comes after. [Grafy Chat](/chat) lets you put the same prompt to many models side by side, so a ranking like this one can be checked against your own task before you change anything. For the agentic and terminal-style evaluations that feed the index, [Grafy Code](/code) is where those differences would show up first — an AI coding studio that installs packages and runs what it writes. If the 1M-token context is what caught your attention, [Reader](/reader) is the natural place to test it, since reading, searching and questioning long PDFs and papers is exactly the workload that context size is meant to serve. And if cost per task is your constraint rather than raw intelligence, the Grafy Blog's guide to [lowest-cost model routing](/blog/lowest-cost-model-routing) describes how to spend the cheapest model that can actually answer. None of this requires the new DeepSeek release specifically; the point is that a model's published rank and its behaviour on your documents are two different measurements, and only one of them is yours. ## Frequently asked questions ### What is DeepSeek V4.1 Flash 1017 (Max)? It is an open-weights model released in October 2026, according to Artificial Analysis. The page lists 552B total parameters with 16B active per token, text and image input, text output, a 1M-token context window, an MIT license and weights on Hugging Face. It scores 45 on the Artificial Analysis Intelligence Index, ranking second of 121 models. ### How much does it cost to run? Artificial Analysis lists $0.30 per 1M input tokens and $1.20 per 1M output tokens, with a 98% cache discount, describing both as moderately priced against medians of $0.43 and $1.68. Weighted across the Intelligence Index evaluations, the page reports an average cost of $0.26 per task. Because the weights are MIT-licensed, self-hosting is also an option.

How fast is it, and how much does it write?

The page ranks it #6 of 79 on speed at 230.4 output tokens per second, which it calls notably fast. It is far less restrained on length: it generated 210M output tokens on the Intelligence Index against a median of 140M, ranking #39 of 47 on verbosity, so longer answers partly offset the low per-token price.

---

Source: Hacker News — front page — DeepSeek V4.1 Flash 1017 (Max)