DeepSeek V4.1-Flash-1017 release roundup: gains on all 11 benchmarks, 77.0 on DeepSWE
A roundup of the launch: web, app and API availability, MIT weights on Hugging Face, 2.0–11.1 point gains across 11 benchmarks, 77.0 on DeepSWE ahead of GLM-5.3 and Opus 5, plus tuning for Harness and Tencent CodeBuddy.
DeepSeek just released V4.1-Flash-1017, a faster and cheaper update to V4.1 Flash. Here's what you need to know. DeepSeek posted at 3:35 PM on Oct 17, 2026 that DeepSeek-V4.1-Flash-1017 is live on its website, app and API. The post pulled about 652K views. DeepSeek says it has stronger benchmark performance, faster task completion and shorter outputs than the previous version, a Pareto improvement in performance and cost. This comes a day after OpenCode made GLM 5.3 Flash free, two days after it did the same for DeepSeek V4.1 Flash, and after the long DeepSeek outage on Oct 15. The model is open source. Weights are on Hugging Face and ModelScope. OpenRouter has it live too, with DeepSeek as the first provider and more on the way. A model named deepseek-v4.1-flash-1017 briefly showed in DeepSeek's API model list the night before, then was rejected again. It was tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek Harness 0.2.1 supports it, and DeepSeek recommends v0.2.1-rc.1 or later. DeepSeek says it was optimized for Harness during training, which cuts task runtime, output usage and token costs. It uses reinforcement learning with self-feedback to build full projects from scratch, from games to 3D modeling. On benchmarks, it beats the previous V4.1 Flash on all 11 benchmarks DeepSeek reported, by 2.0 to 11.1 points. Its best is 77.0 on DeepSWE v1.1. TestingCatalog says that tops GLM-5.3, Kimi K3, GPT-5.6 Sol and Opus 5. Users say it beats Opus 5 on DeepSWE, Agents Last Exam and Automation Bench, and call it the top open-weight model. Hands-on coding tests reportedly show a gap between benchmark scores and real output. Search and Fetch APIs are now open to everyone too. They run on DeepSeek's in-house search engine, which serves hundreds of millions of requests a day. Key numbers: - API price off-peak: $0.15 per 1M input tokens, $0.60 per 1M output tokens - API price at peak: $0.30 input, $1.20 output - Cache read: $0.003 per 1M tokens - DeepSWE v1.1 score: 77.0 - Gains over old Flash: 2.0 to 11.1 points on 11 benchmarks - Search and Fetch: $0.75 per 1,000 calls each DeepSeek-V4.1-Flash-1017 is a minor update to V4.1 Flash, a sparse mixture-of-experts model.
This post quotes another post: x.com/deepseek_ai/status/2111375640717033777