DeepSeek V4.1 Flash 1017
A minor version iteration of the large language model in DeepSeek's Flash series, with targeted enhancements in reasoning, coding, agentic tasks and image understanding. It supports repository-level code editing and terminal operations, and produces shorter and faster output.
Core Facts
Compared with the previous version, DeepSeek V4.1 Flash 1017 completes tasks faster and produces shorter output.
DeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, surpassing Claude Opus 5's 74.0.
V4.1 Flash surpassed Opus 5 on the DeepSWE benchmark by a narrow margin of 0.2 points.
Timeline (last 90 days)
Testing by a third-party inference service team shows that V4.1 Flash 1017 mainly wins by producing shorter outputs: the median output for math competition problems dropped from 1681 tokens to 1124, and the longest reasoning chain dropped from 3814 to 1880, while accuracy showed no significant change.
The official cost comparison chart shows that the new model's highest point costs a bit over 2 yuan (at the 77% position), while OPUS-5 is close to 100 yuan (around 70%).
Among the 11 benchmarks officially published, the terminal operation test version 4.0 rose from 31.2 to 42.3, the deep software engineering test version 1.1 rose from 74.2 to 77.0, the automation test reached 60.5, and the Agent ultimate exam scored 35.4.
V4.1 Flash 1017's lowest reasoning tier outputs 790 tokens, a 60% reduction from nearly 2000 tokens in the old version, and it can still answer questions correctly.
Official pricing is unchanged: output at 8 yuan per million tokens, half price during off-peak hours (the 2 a.m. slot), with open-source weights allowed for commercial use; the search and crawl interface costs 5 yuan per 1000 calls, billed only on success, with 5 calls per second per interface and 10 results returned by default.
In the comparison with OPUS-5, V4.1 Flash 1017 won 4 of 11 items and lost 7, with 42.6 versus 56.3 on the Humanity's Last Exam and 42.3 versus 51.8 on terminal operation.
DeepSeek V4.1 Flash 1017 launched on October 17, only 37 days after the previous version (released on September 10).
Someone posted in a community that using Model Trace to run a fingerprint test on DeepSeek V4.1 Flash 1017, the result was identified as GPT-65 AllStar with a probability as high as 90.71%.
The Bilibili video is titled "DeepSeek V4.1 Flash 1017 vs Gemini 4 Argon," but its actual content is an interpretation of Hayek's "The Road to Serfdom," so there is a mismatch between title and content.
DeepSeek 4.1 Flash and Gemini 4 Argon are two large models released in the same period.
40 more timeline events
All Facts (20)
Compared with the previous version, DeepSeek V4.1 Flash 1017 completes tasks faster and produces shorter output.
75%VerifiedDeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, surpassing Claude Opus 5's 74.0.
70%VerifiedV4.1 Flash surpassed Opus 5 on the DeepSWE benchmark by a narrow margin of 0.2 points.
70%VerifiedDeepSeek v4.1 Flash 1017 has been released and has sparked widespread attention in technical communities, but official details disclosed are limited.
80%VerifiedV4.1-Flash-1017 scores 77.0 on the DeepSWE benchmark.
65%VerifiedAccording to AI leaker account Lumina, DeepSeek V4.1 Flash has launched a Fast version, but there is no official announcement or pricing yet.
65%VerifiedThis update adds support for the official model ID deepseek-flash; new sessions default to V4.1 Flash, but explicitly configured models take higher priority.
65%VerifiedArtificial Analysis evaluation shows DeepSeek V4.1 Flash scored 40 on the Intelligence Index, surpassing V4 Pro to become DeepSeek's new flagship, featuring native multimodal input, 8B activated parameters on input and 16B on output, released under the MIT license.
65%VerifiedV4.1 Flash activates 8 billion parameters when processing input and activates 16 billion parameters when generating output.
65%VerifiedDeepSeek V4.1 Flash adopts a Causal Encoder-Decoder asymmetric architecture with a total parameter count of 552B.
65%VerifiedDeepSeek V4.1 Flash was released on September 10 with MIT open-source weights.
65%VerifiedDeepSeek released DeepSeek-V4.1-Flash, a method that significantly compresses the memory required for KV cache.
70%Partially VerifiedDeepSeek V4.1-Flash-1017 improves by 11 points on the Terminal-Bench 4.0 benchmark compared with the previous version.
75%Partially VerifiedIn a Minecraft sandbox game generation comparison, Kimi K3 took 2.5 hours and cost $35, while DeepSeek V4.1 Flash 1017 took 1.5 hours and cost 80 cents, nearly 44 times cheaper.
70%Partially VerifiedIn the World of AI Benchmark multi-domain comprehensive test, V4.1 Flash 1017 currently ranks 10th or 11th.
65%Partially VerifiedMaximum output tokens: MiniMax M3.1 Flash is 512K, DeepSeek V4.1 Flash 1017 is 384K, and Claude Haiku 5.5 is 128K.
65%Partially VerifiedCalculated per thousand repair tasks, Claude Haiku 5.5 costs about $0.70, DeepSeek V4.1 Flash 1017 about $1.35, and MiniMax M3.1 Flash about $4.45 by default and about $0.64 at the low setting.
65%Partially VerifiedV4.1 Flash 1017's score on Terminal-Bench 4.0 rose from 31.2% to 42.3%, a relative improvement of 35% in five weeks.
65%Partially VerifiedIn the "pelican riding a bicycle" SVG prompt test, DeepSeek V4.1 Flash 1017's output cost 23 cents, while Opus 5.5 cost $6, making the former about 26 times cheaper, but DeepSeek took 1 hour while Opus took only 20 minutes.
65%Partially VerifiedIn the cross-domain tests of the World of AI Benchmark, DeepSeek V4.1 Flash 1017 ranks 10th to 11th overall.
65%