DeepSeek V4.1-Flash-1017 AI Model Delivers 75% Hardware Savings But Struggles With Hallucinations
9 Sources
[1]
Deepseek's latest AI model is smarter than ever, but its low running costs come at a price -- uses only one-quarter of the HBM and one-eighth of SSD storage but still battles hallucinations in agentic workloads
Frontier models are far more capable but paying a premium is quickly becoming a matter of option Deepseek just upgraded its V4.1 Flash open-weight model to revision 1017. What looks like a mere point revision may actually be a significant step forward in model efficiency. On the product page,
[2]
DeepSeek releases V4.1-Flash-1017 AI model with better scores
DeepSeek opened its in-house search and fetch APIs to all users at $0.75 per 1,000 calls for each endpoint DeepSeek released an updated version of its V4.1-Flash AI model on Saturday that the company said finishes tasks more quickly and writes more concise responses than its predecessor. The
[3]
DeepSeek updates V4.1-Flash model to improve performance
According to DeepSeek, the updated model has 'substantial improvements' across reasoning, coding, agentic tasks and visual benchmarks. Chinese AI company DeepSeek has released an update of its V4.1-Flash model, which it said has stronger performance, faster task completion and shorter outputs than
[4]
DeepSeek releases updated V4.1-Flash model it says completes tasks faster
Chinese artificial intelligence startup Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. has released an updated version of its V4.1-Flash model that it says completes tasks faster. DeepSeek-V4.1-Flash-1017 arrives a little over five weeks after the company launched the
[5]
DeepSeek updates V4.1 Flash without raising API prices
The updated model scored 77% on DeepSWE v1.1 in DeepSeek's tests, up from 74.2% for the September version, although the company did not quantify its speed improvements. Chinese artificial intelligence startup DeepSeek released DeepSeek-V4.1-Flash-1017, an updated model it says completes tasks
[6]
DeepSeek AI Model Update: DeepSeek pushes out updated AI model it says is faster, leaner
DeepSeek released an update of its AI model this weekend that it says is faster and leaner, ramping up the Chinese company's challenge to Silicon Valley rivals Anthropic PBC and OpenAI as well as domestic competitors like Moonshot Kimi and Z.ai The firm unveiled the V4.1-Flash-1017 late Saturday
[7]
New DeepSeek V4.1 Flash 1017 Hits 300 Tokens per Second in Tests
DeepSeek V4.1 Flash 1017 has emerged as a noteworthy addition to the open source AI landscape, offering a balance of affordability and functionality that appeals to a wide audience. In a detailed breakdown by World of AI, the model's performance upgrades are highlighted, including its ability to
[8]
DeepSeek Deploys What It Calls Faster Version of Its Open-Source Model | PYMNTS.com
China's DeepSeek has unveiled what it calls a faster and leaner version of its artificial intelligence model. The company describes the DeepSeek-V4.1-Flash-1017 model, announced Saturday (Oct. 17) as offering "[s]tronger benchmark performance, faster task completion, and shorter outputs than the
[9]
DeepSeek releases AI model update, opens search APIs to all users By Investing.com
Investing.com-- Chinese artificial intelligence startup DeepSeek on Saturday released an updated version of its AI model, touting improved benchmark performance, faster task completion, and shorter outputs as it seeks to enhance efficiency and broaden access to its tools. The company said
Share
Copy Link
DeepSeek released V4.1-Flash-1017, an updated open-weight AI model that uses one-quarter of the HBM and one-eighth of SSD storage compared to its predecessor. The model scored 77% on DeepSWE v1.1 software engineering benchmark, outperforming Claude Opus 5's 74%, but hallucinated 67% of non-answers in reliability tests.
DeepSeek V4.1-Flash-1017 Achieves Major Hardware Efficiency Gains
DeepSeek launched V4.1-Flash-1017 on October 17, 2026, marking a significant leap in AI model efficiency despite being labeled a minor update
1
. The open-weight AI model delivers substantial hardware savings, using only one-quarter of the HBM and one-eighth of SSD storage for its KV cache compared to V4 Flash, spending just 890 bytes per token entry1
. This Mixture-of-Experts model contains 552B total parameters with 16B active during runs and supports multimodal inputs with a 1 million token context window1
. The model is now live on DeepSeek's website, app, and API, with weights available on Hugging Face and ModelScope under the MIT license4
.
Source: Geeky Gadgets
Strong Performance on Software Engineering Benchmark DeepSWE
The updated model achieved a 77% task completion rate on DeepSWE v1.1, improving from the September version's 74.2%
2
3
. This performance surpasses both GPT-5.6 Sol at 73% and Claude Opus 5 at 74-75%1
3
. Additional benchmark scores demonstrate substantial improvements across multiple areas: Humanity's Last Exam rose to 42.6 from 36.8, Terminal-Bench 4.0 increased to 42.3 from 31.2, and ProgramBench climbed to 29.1 from 20.32
. DeepSeek optimized the model specifically for agentic workloads, including DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, recommending users upgrade to DeepSeek Harness 0.2.1-rc.1 or later2
5
.API Pricing Remains Unchanged at Industry-Low Rates
DeepSeek maintained its aggressive API pricing strategy, charging $0.15 per million input tokens and $0.6 per million output tokens during off-peak times
1
. Cache hits reduce input token costs to just $0.0031
. The company opened its Search & Fetch APIs to all users at $0.75 per 1,000 calls for each endpoint, undercutting competitors like Parallel at $11
2
. Calls made through DeepSeek Harness, Claude Code, and OpenAI's Codex remain free until October 31, 20262
5
. When the original V4.1-Flash launched in September with these prices, shares of Chinese rivals MiniMax Group Inc. and Alibaba Group Holding Ltd. fell in Hong Kong4
5
.Related Stories
Hallucination Concerns Limit Reliability for Open-Ended Tasks
Despite strong benchmark scores, V4.1-Flash-1017 struggles with hallucinations, answering questions correctly only 43% of the time on Artificial Analysis' Omniscience benchmark
1
. More concerning, the model hallucinated 67% of non-answers, earning just 5 points on the main index compared to Kimi K3-max's 20 and frontier models' 42+1
. This reliability gap means the model excels at tasks with verifiable outputs like programming, OCR with review, and format conversions where verification steps can catch errors, but struggles with open-ended research, fact-checking, and drafting precise documents1
. The trade-off reflects a broader pattern where open-weight AI models exchange reliability for dramatically lower running costs, particularly for long-running agentic tasks1
.
Source: Silicon Republic
Enhanced Capabilities in Game Development and 3D Modeling
DeepSeek incorporated reinforcement learning with self-feedback to improve end-to-end project development capabilities
2
3
. The company showcased these enhancements through examples including a 3D model of China's Forbidden City, a web-based racing game, and a recreation of Portal 2 built with the open-source Godot engine3
. The company markets the model with the slogan "wish-driven programming made real," referencing colloquial "vibe coding" to signal its fitness for open-ended tasks with broad descriptions1
. The release intensifies competition between DeepSeek and rivals including Anthropic, OpenAI, and Moonshot Kimi in the global race for AI dominance2
. As of early October 2026, DeepSeek was raising at least 80 billion yuan ($12 billion) at a minimum valuation of 500 billion yuan, with Tencent Holdings and Contemporary Amperex Technology Co. among top investors2
. The update follows a nearly five-hour API outage on October 15 that experts attributed to insufficient computing capacity, after which DeepSeek apologized and committed to removing single points of dependency4
5
.
Source: Tom's Hardware
References
Summarized by
Navi
[3]
[5]
Related Stories
DeepSeek V4 closes gap with frontier models while slashing AI costs by 75%
24 Apr 2026•Technology
DeepSeek V4-Flash Emerges as Cheapest Major AI Model at $0.28 Per Million Tokens
04 Aug 2026•Technology
DeepSeek Launches Multimodal AI Model Approaching Anthropic Opus 4.8 Performance at Fraction of Cost
22 Aug 2026•Technology
Recent Highlights
Recent Highlights
Today's Top Stories
Don’t drown in AI news. We cut through the noise - filtering, ranking and summarizing the most important AI news, breakthroughs and research daily. Follow topics that matter to you and stay ahead.
