DeepSeek Open-Sources Inference Framework, Challenges OpenAI, Anthropic

Key Takeaways

• <ul><li>DeepSeek open-sourced DeepInfer under Apache 2.0, powering DeepSeek-V4.1-Flash with support for NVIDIA Blackwell GPUs and Huawei Ascend 950 NPUs</li><li>DeepSeek-V4.1-Flash is priced at $0.035 per million input tokens and $0.12 per million output tokens, undercutting OpenAI GPT-6 Astra</li><li>One DeepInfer cluster can serve 155 trillion tokens per day with decode speeds above 25,000 tokens per second per GPU</li><li>F5 reports 78% of organizations now run AI inference across an average of seven models, but only 2% were equipped for scale a year ago</li><li>IDENTOS PolicyArc is now generally available, offering task-specific, time-limited access for AI agents with automatic revocation and audit trails</li><li>The Arkansas State University System Board approved four IT policies including an AI Governance Policy and Data Security and Classification Policy</li><li>A new study reveals factor coverage bias in LLMs used for clinical decisions, risking premature conclusions and unnecessary medical tests</li><li>The IMF warns AI risks including job displacement, inequality, and cybersecurity threats may outweigh benefits globally</li><li>Robinhood experts say human control remains vital in AI-driven investing to maintain skepticism over automated decisions</li><li>Vendavo's AI Deal Desk Assistant claims to boost seller productivity by up to 3 times and improve gross margins by 100 to 300 basis points</li></ul>

DeepSeek open-sources DeepInfer inference framework

DeepSeek released DeepInfer, its internal inference framework, as open source on GitHub. The system powers DeepSeek-V4.1-Flash in production and supports high throughput with low latency. It includes high-performance kernels, distributed inference, speculative decoding, and structured output support. DeepSeek aims to help the community lower compute costs, reduce energy use, and make large-scale AI deployment more affordable.

DeepSeek makes inference tech open source to cut AI costs

DeepSeek open-sourced DeepInfer, its in-house inference framework, to help make AI more affordable. The framework includes optimized inference engines, distributed inference, speculative decoding, and structured outputs. DeepSeek said it will keep working with the community to improve efficiency for both private deployments and public clouds. The release came one day after the U.S. Department of Homeland Security cited DeepSeek models as a threat vector.

DeepSeek open-sources DeepInfer framework for V4.1-Flash

DeepSeek released DeepInfer, its production inference engine, under an Apache 2.0 license on GitHub. It currently supports DeepSeek-V4.1-Flash and runs on NVIDIA SM100 Blackwell GPUs and Huawei Ascend 950 NPUs. DeepSeek reports one cluster can serve 155 trillion tokens per day, with decode speeds above 25,000 tokens per second per GPU. The release does not yet include online serving, KV cache systems, or load balancing.

DeepSeek challenges OpenAI and Anthropic with open-source tool

DeepSeek released DeepInfer, an open-source inference platform built to run DeepSeek-V4.1-Flash. The company said it wants to help lower compute costs, reduce energy use, and make AI more affordable. DeepSeek-V4.1-Flash is priced at $0.035 per million input tokens and $0.12 per million output tokens, much cheaper than OpenAI GPT-6 Astra. The release adds pressure on American AI companies facing rising inference costs and competition from Chinese open-source models.

Closing the agent governance gap

AI agents are creating security risks inside organizations because companies often lack visibility into what agents are doing. The problem grows as employees can easily build agents using vendor tools, platform features, or homegrown code. Security teams need one view of all agents, including their permissions and behavior, to spot risks that no single platform can see on its own. The article calls for both build-time security checks and runtime monitoring to control what agents can do and prevent unauthorized actions.

AI inference path determines enterprise value and security

F5 reports that 78% of organizations now run AI inference across an average of seven models. Research found just 2% were equipped to operate AI at scale a year ago. The inference path includes retrieval layers, routing logic, identity and policies between user intent and model output. This path determines AI accuracy, cost, performance, resilience and security at scale.

A-State Board approves new AI and technology policies

The Arkansas State University System Board approved four policies to improve IT management and security across campuses. The new policies include an Artificial Intelligence Governance Policy, Data Security and Classification Policy, and Information and Technology Resource Security Policy. System President Todd Shields said implementation procedures will ensure consistency. The board also approved budget increases and a new nonprofit organization.

Study reveals LLM bias in medical decision determinability

A new study identifies factor coverage bias as a failure mode in large language models used for clinical decisions. LLMs often judge determinability based on the amount of decision-relevant information rather than whether missing information could change the conclusion. Researchers built a controlled pairwise benchmark to test this bias. The findings show LLMs may make premature conclusions or gather unnecessary information like invasive tests.

IMF warns AI risks may outweigh benefits globally

The International Monetary Fund warns AI risks could outweigh benefits as global investment surges. The Global Financial Stability Report highlights job displacement, increased inequality and cybersecurity threats. IMF notes AI benefits like productivity gains may not be evenly distributed. The fund urges policymakers to proactively address risks and promote responsible AI development.

Robinhood experts say human control remains vital for AI

Robinhood experts at the 3rd HOOD Summit in Houston say human judgment remains crucial in AI-driven investing. VP of Product Management Abhishek Fatehpuria and Chief Investment Officer Steph Guild emphasized that AI products give investors a fair shot against institutional players. They stressed that humans in the loop can show skepticism over AI decisions. Experts say AI agents are not magic black boxes.

Vendavo launches AI assistant for sales teams

Vendavo released the AI Deal Desk Assistant inside its Quoting software. The tool gives sales teams customer and pricing data in plain language. It can create draft quotes and handle approvals automatically. Vendavo says it can boost seller productivity by up to 3 times and improve gross margins by 100 to 300 basis points.

Experts warn AI could make people less capable

Experts are warning that heavy reliance on AI tools could lead to cognitive surrender. People may lose important skills because AI does tasks for them. Researchers say this disempowerment is a serious risk. They note that if systems fail, there may be no human available to fix problems. The concern grows as AI takes on more consequential decisions.

IDENTOS PolicyArc now available for AI agent access

IDENTOS announced that PolicyArc is generally available. The tool lets security teams set task-specific and time-limited access for AI agents. Access is automatically revoked when a task ends. It checks consent and real-time context before granting permissions. PolicyArc also keeps a detailed audit trail of every access decision.

Biochemistry student frustrated with AI grading

A biochemistry major at Miami University says AI grading on homework feels unfair. The platform Stemble uses AI to grade written responses. The student says the AI demands specific wording and extra details not asked for. This forces students to spend time rewriting instead of learning. The student believes AI should not grade student work.

Sources

• DeepSeek Open-Sources DeepInfer, Its In-House Inference Framework
• DeepSeek Open Sources Inference Tech to Lower AI Costs
• DeepSeek Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash
• DeepSeek Dials Up the Pressure on OpenAI and Anthropic With Open-Source Inference Tool
• Closing the agent governance gap
• The inference path is where AI value is won or lost
• A-State Board approves AI, technology and security policy changes
• When to Stop Asking: Factor Coverage Bias and LLM Determinability Judgments with Underspecified Information in Medicine
• IMF warns AI risks could outweigh benefits as global investment surges
• Human Touch Still Needed to Power the AI ‘Magic Money Box,’ Say Robinhood Experts
• Vendavo Launches AI Deal Desk Assistant, Bringing Deal Intelligence Directly into the Quoting Workflow
• Experts Warn of Humans Becoming Empty Vessels Controlled by AI
• IDENTOS PolicyArc gives AI agents task-specific, time-limited access
• As a biochemistry major, I hate using AI in my classes