DeepSeek V4.1 Flash 1017

DeepSeek-V4.1-Flash was released on September 10, 2026, and the Coursiv blog lays out the architectural changes versus V4 Flash: backbone parameters grew from 284B to 552B, active parameters per token fell from 13B to 8B for reading input and 16B for generating output, the architecture shifted from a Mixture-of-Experts The post also reports Terminal-Bench 2.1 rising from 82.7 to 90.6, Terminal-Bench 4.0 from 7.0 to 31.2, and DeepSWE v1.1 from 54.4 to 74.2 versus V4 Flash. From 12:00 Beijing time on September 14, 2026 (04:00 UTC), every request to `deepseek-v4-pro` is routed to V4.1 Flash and billed at Flash rates, a cut of roughly 77

DeepSeek released V4.1 Flash on 10 September 2026 and the third-party write-up cites the official model card reporting a 552B-parameter Mixture-of-Experts model that activates 8B parameters per token during prefill, supports a 1M-token context window, reads images natively, and ships under the MIT license with $0.30 pe The article reports benchmark comparisons versus GPT-5.6 Sol and Claude Opus-5.0 drawn from the official model card on DeepSWE v1.1 (74.2 vs 73.0), AutomationBench (54.8 vs 45.8), Agent's Last Exam (31.8 vs 26.7), and CyberGym (88.1 vs 84.5). All figures are presented as restatements of DeepSeek's own published model c

DeepSeek released DeepSeek-V4.1-Flash-1017, stating it delivers stronger benchmark performance, faster task completion, and shorter outputs than the previous version — a claimed Pareto improvement in performance and cost. The model is live on DeepSeek's website, mobile app, and API, and DeepSeek publishes a DeepSWE v1.1 Pass@1 versus cost-per-task chart measured with the mini-swe-v2 harness. The new checkpoint was trained with reinforcement learning using self-feedback to improve end-to-end project development, with demos covering 3D modeling of the Forbidden City, a web-based racing game, a Godot-engine Portal 2 recreation, and a prehistoric-to-modern 3D scene. It is optimized for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy, and WorkBuddy, with open-source weights published on Hugging Face and ModelScope.

The BenchLM model profile records DeepSeek V4.1 Flash as released on 10 September 2026 with open weights, reasoning capability, and a 1M-token context window. It documents text and image input modalities, an API model ID of `deepseek-flash`, and lists 22 sourced benchmark rows across agentic (8/8 verified), coding (6/6 Stated pricing on the page is $0.30 input per million tokens, $1.20 output per million tokens, $0.006 cached, with a blended rate of $0.75 and a reported speed of 212 tok/s with a first-token latency of 10.62 s. All cited figures link back to DeepSeek's own docs, pricing, and vision guide, and no OpenCode Zen serving-p

DeepSeek officially released DeepSeek-V4.1-Flash on 2026-09-10 as the smallest model in its new architecture family, with native multimodal visual understanding, a higher capability ceiling, faster inference, and higher throughput. The DeepSeek API changelog lists benchmarks including GPQA Diamond 90.9, HLE 36.8 (39.1 The API change renames the call target to `deepseek-flash` for V4.1 Flash, while previous-generation V4 Flash and V4 Flash Vision Exp were retired with their old model IDs (`deepseek-v4-flash`, `deepseek-v4-flash-vision-exp`) temporarily routed to V4.1 Flash for compatibility. DeepSeek V4 Pro API service continues past