Intel / post

DeepSeek V4.1 Flash 1017 jumps six points and ties for #2 open weights

Source: x.com
Date: Published Oct 19, 2026

Context

DeepSeek has shipped an updated checkpoint of its V4.1 Flash model, labelled 1017, and Artificial Analysis reports that it scores 45 on its Intelligence Index, a composite of the evaluations it runs, six points above the original V4.1 Flash (Artificial Analysis on X). The price and did not move: the update keeps the 1M- context window and the first-party price of $0.30 per million input tokens and $1.20 per million output tokens at peak, with off-peak rates half that. Artificial Analysis lists it at 552B total parameters with 16B active per token, under the MIT license. The score change is therefore a capability gain at the same per-token price. Artificial Analysis's earlier assessment of the original V4.1 Flash is the baseline it is measured against.

The biggest single movement is on AA-Omniscience, a knowledge test that, by the firm's methodology, credits correct answers, subtracts for wrong ones, and does not penalize a model for declining to answer. The original V4.1 Flash almost never held back: its non- rate was 4%, which rose to 33% in the update, while accuracy edged up from 43% to 46%. Because wrong answers cost points, the overall AA-Omniscience Index went from -5, meaning more wrong answers than right ones, to +5. Artificial Analysis also reports Elo gains of 146 points on GDPval-AA v2.1 (to 1746) and 150 points on AA-Briefcase v1.1 (to 1570), two evaluations it groups under agentic capability, and a rise from 14% to 24% on the CritPt scientific-reasoning test.

On ranking, Artificial Analysis places the update level with GLM-5.3 as the second-highest open-weights model, one point behind MiMo-V2.6-Pro at 46 and ahead of Step 5 Preview and Kimi K3 at 44. On cost, it puts the model's cost per task at $0.26, compared with GLM-5.3-Flash at $0.28 for a score of 42. The same comparison notes that MiMo-V2.6-Pro, at $0.13 per task, and GPT-6.1 Sol on medium effort, scoring 48 at $0.21, both deliver more capability for less, so the update improves DeepSeek's position without making it the cheapest route to that level of performance.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Read the source x.com