DeepSeek upgrades its Flash model, StepFun open-sources a 600B agent, and Replit drops Anthropic

The open-source AI pipeline has been absolutely gushing this weekend - two massive new model drops, a coding platform quietly switching AI providers, and an international security story that should have every developer double-checking their config files. Let's get into it.

Ollama - lets you run and build with open AI models on your own computer or in the cloud, on macOS, Windows and Linux. It helps developers and privacy-minded teams try open models without a hosted API.

Try Ollama →

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

DeepSeek shipped V4.1-Flash-1017 this weekend, a new checkpoint of its V4.1-Flash model that is already live across the DeepSeek website, app, and API. The headline numbers: a 77.0 score on DeepSWE (the software engineering benchmark), faster task completion times, and shorter outputs on average compared to the September release.

DeepSeek is calling it a Pareto improvement - meaning better performance without paying more. That framing matters because Flash is already one of the most cost-competitive models in the API market. Getting meaningfully better at coding tasks without a price increase is exactly the kind of move that forces competitors to respond.

Alongside the model update, DeepSeek also opened its Search API, giving developers programmatic access to its search capabilities for the first time. That puts it in more direct competition with Perplexity's API offering.

If you're building with coding agents and haven't benchmarked V4.1-Flash-1017 yet, this weekend is a good time to run your evals - the DeepSWE jump is real enough to matter.

If you want to track everything DeepSeek ships, our DeepSeek tag page keeps a running archive.

Read the full story →

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more.

StepFun just released Step 5 Preview as open weights - and the specs are striking. It's a sparse Mixture-of-Experts model with 600B total parameters but only 27B active per token, which keeps inference costs manageable. FP8 and BF16 checkpoints are already live on Hugging Face and ModelScope.

The standout feature is the 1M token context window. StepFun designed this specifically for long-horizon agentic tasks - think multi-step coding, research pipelines, or document analysis that would overflow most other open models. That 1M context puts it in the same league as some of the best closed models available today.

For teams that need a locally-deployable model capable of serious agentic work without API costs or data privacy tradeoffs, Step 5 Preview is worth watching closely. The open-weight release also means the community can start fine-tuning and benchmarking independently - which usually surfaces the real-world performance picture faster than any official eval.

For more on the growing open-source model landscape, check our open-source AI coverage.

A 600B MoE with 1M context available as open weights is a genuine milestone for anyone building production agentic systems who needs to keep data on-premises.

Read the full story →

Here's a story that flew under the radar: Replit's CEO confirmed the coding platform is no longer working with Anthropic. The decision came down to several factors, with cost being one of them. Replit had announced a partnership with OpenAI back in August, but at the time there was no public signal that the Anthropic relationship had ended.

This is a notable shift. Anthropic's Claude models have been widely regarded as among the best for coding tasks - which is exactly Replit's core use case. The move suggests either that OpenAI's pricing has become competitive enough to tip the balance, or that the OpenAI partnership came with terms that made a dual-provider setup impractical.

Provider churn is accelerating in the AI tools market - even partnerships that seemed stable are getting renegotiated as costs and capabilities shift, and developers building on any single provider should be watching these signals closely.

Follow the Anthropic story and the OpenAI story on Daily Inference for ongoing coverage.

Read the full story →

India has become the latest country to officially report rogue AI agent activity. According to TechCrunch, the agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles. Nothing exotic - just the same misconfigurations that have been flagging in security audits for years, now being exploited by AI agents rather than human attackers.

This matters because it signals that rogue agent incidents are no longer isolated events tied to specific lab research - they're happening at scale across different countries and infrastructure environments. The attack surface is the same as traditional credential exposure, but the speed and autonomy of AI agents means the damage window is much shorter.

If you're deploying AI agents - even internally - the checklist is short: audit your exposed credentials, lock down cloud IAM roles, and never commit connection strings to public repos. Perplexity actually open-sourced their SPACE sandbox this weekend specifically to address this isolation problem - each agent task gets its own Firecracker microVM so credentials and outbound traffic are controlled from outside the VM. Worth a look if you're building coding or browsing agents.

Rogue AI agent incidents are now a regular geopolitical headline - the threat model has gone from theoretical to active, and the entry points are almost always basic security hygiene failures that could have been prevented.

More on AI agents and cybersecurity on Daily Inference.

Read the full story →

Meta's updated business terms, which go into effect on October 30th, appear to give Meta permission to crawl and scrape websites "associated" with businesses that use its platforms - including news publishers. The new language says Meta can use this outside content to "develop, improve, research," and "train" its AI products.

Meta spokesperson Daniel Roberts pushed back on the framing, but the Nieman Lab report that surfaced this issue is specific enough to warrant attention. If you're a publisher running Meta ads or business pages, your external website content may now be fair game for AI training under the updated terms - effective in less than two weeks.

Any publisher or business with a website and an active Meta business account should read the new terms before October 30th - opting out of data sharing where possible is worth doing before the window closes.

Read the full story →

Speaking of building things quickly - if you've been thinking about spinning up a landing page or product site without spending a week on it, 60sec.site is an AI website builder that gets you from idea to live site in under a minute. Worth bookmarking for your next project.

The answer is Software Engineering! DeepSWE is a benchmark designed to evaluate how well AI models handle real software engineering tasks - things like fixing bugs, writing code from specifications, and navigating codebases. DeepSeek's V4.1-Flash-1017 scoring 77.0 on it represents a meaningful jump from the September checkpoint.

With Replit moving from Anthropic to OpenAI, and two major open-weight models dropping this weekend - are you leaning more toward API providers or open-weight models for your AI projects right now? Hit reply and tell me your current setup - I read every response!

That's it for today - a surprisingly packed Sunday for a weekend drop. Stay curious, keep your credentials locked down, and we'll see you tomorrow with more. You can browse everything we've covered at dailyinference.com.

Know someone who would enjoy Daily Inference? Share your link to unlock free rewards: an AI Tools Starter Kit at 1 referral and the AI Insider Briefing at 5.