Mercury 3 lands cheap, MiniMax M3.1 Flash goes free, and the open model pile grows

What shipped

On 19 October, the day's releases clustered around cost: Inception's Mercury 3 and MiniMax's M3.1 Flash both landed on routing services at launch or free pricing, while Hugging Face's trending board filled with open vision and text models shipped in compressed formats. Underneath sat harder news, from an expected shakeout among AI cloud providers to Microsoft calling employee-used agents an insider risk.

Hugging Face trending

The trending board leaned toward size plus compression: three of the five models moving up the Hub read images as well as text, and two arrived in reduced-precision formats so they run on rented or owned hardware rather than only in a lab. Nvidia, Meta's model arm, and the MiniCPM team are all betting that open vision models plus cheaper weights beat closed APIs on price for document, screen, and photo work.

• •Beam-1.0-501B-A23B reflection's text model tops the Hub board at 501 billion parameters with 23 billion active per token, a mixture-of-experts design that keeps per-token cost below what the raw size suggests. Vibe builders can pull it for fine-tuning and inference.
• •Muse-Spark-Open-816B-A42B-NVFP4 meta-models posted a vision-language model at 816 billion parameters, 42 billion active, shipped in NVFP4 (a 4-bit weight format that shrinks memory needs). It suits prompts that mix a screenshot with a question.
• •Timestral-209M mistralai's 209 million parameter model forecasts numeric sequences rather than chat, running on the timestral library (a forecasting toolkit). SMB owners can point it at sales or stock histories without hiring a data team.
• •NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 Nvidia's vision-language model reads images and text at 120 billion parameters with 12 billion active, built on transformers (the standard Python toolkit for running models). It targets document and screen understanding.
• •MiniCPM-V-4.7-35B-A3B-gguf openbmb packaged its 35 billion parameter vision model in gguf (a packed file format for llama.cpp, a lightweight local runtime), so image understanding can run on a single GPU (graphics processing unit) instead of a rented cluster.

Vendor launches

Vercel carried the bulk of the day with three items: rewritten Hobby and Pro terms of service, plus two models added to its AI Gateway routing layer, Inception's Mercury 3 at 50 percent off and MiniMax's M3.1 Flash free through 20 October. Nvidia supplied the other two, one with the German government on industrial AI and one with the European Broadcasting Union on security for media agents.

• •Vercel terms of service update Vercel rewrote Hobby and Pro terms effective 19 October, adding a section explaining that plan-included credits expire at the end of the subscription term, plus cancellation and service-change rights for eligible UK and EU consumers. Check your billing terms before scaling.
• •Nvidia and Germany German Chancellor Friedrich Merz and Nvidia chief Jensen Huang met at Berlin's Kraftwerk venue to map industrial AI adoption, linking Nvidia's hardware to German manufacturing and automotive deployments. Industrial buyers, not labs, are the growth target.
• •Mercury 3 on AI Gateway Inception's Mercury 3 arrived on Vercel's AI Gateway at 50 percent off launch pricing, generating above 1,000 tokens per second with a 1M-token context window. Its diffusion approach (parallel refinement of text instead of one token at a time) suits browser agents and support bots.
• •European Broadcasting Union and Nvidia The European Broadcasting Union (EBU) is working with Nvidia to extend its sovereign content framework with cloud and open models, aimed at securing agentic workflows for public service broadcasters. Media buyers now treat agent access as a security item.
• •MiniMax M3.1 Flash free on AI Gateway MiniMax's open-weight multimodal model is free on AI Gateway through 20 October, with a 1M-token context window for long documents and multi-step agent sessions. This is a cheap window to test coding and computer-use workflows.

Industry news

Two threads ran through 19 October: money tightening around AI infrastructure, and security catching up to agents that act on their own. Executives expect only seven to ten of hundreds of AI cloud providers to survive, while Microsoft told enterprises to treat employee-used agents like insider threats and Amazon shipped a tool aimed at breaking Nvidia's grip on the PyTorch (the free toolkit most AI apps are built with) stack.

• •Neocloud shakeout forecast Industry executives expect only seven to ten of hundreds of AI cloud providers to survive. CoreWeave lost a quarter of its market value, Microsoft pulled out of a 23,000-chip deal with Nscale, and Nvidia draws 45 percent of data center revenue from buyers outside the hyperscalers.
• •AI used to de-bloat smart TVs Owners are pointing AI agents at smart TVs to strip ads, tracking, and sluggish menus, turning expensive screens back into plain displays. The lesson for builders is that agentic cleanup of consumer software is a paid job, not a demo.
• •Laplace Intelligence The University of Michigan spinout runs AI agents inside interactive simulation to red-team robots and self-driving cars before they reach the road, and says it is already profitable. Testing-as-a-service for physical AI is becoming a category.
• •Goodfire researcher's neural network claim Nick Cammarata's 26,000-word post arguing the field misreads how neural networks work set off wide speculation and celebration. The debate matters practically, because interpreting what a model learned is how teams decide whether to trust it.
• •Tolan raises $30 million The AI companion app passed 250,000 active users before closing a $30 million round. Companion apps keep pulling capital despite open questions about retention, dependency, and safety.
• •Wabi from Replika's founder Eugenia Kuyda launched Wabi, a free personal AI assistant downloadable today, reviving an idea she pitched more than a decade ago. Free assistants now compete directly with paid subscription products.
• •OpenAI engineer custody ruling A judge kept Abdulrahman Alfozan in custody, comparing a Saudi government offer to secure his release to a blank check. Talent and trade-secret disputes in AI now carry geopolitical weight.
• •MiniMax price pressure MiniMax's M3.1 Flash rose 13 points over its predecessor to 42 on the Artificial Analysis Intelligence Index at half the cost, and hit 73.7 percent on the DeepSWE v1.1 coding benchmark for under $0.30 per task. Coding and agent work keep getting cheaper.
• •Microsoft on agents as insider threats Microsoft is urging enterprises to watch the agents employees connect to systems outside labs like OpenAI and Anthropic, calling frontier models untrustworthy. Security teams now audit agent permissions the way they audit user accounts.
• •Amazon TorchNeuron Amazon released TorchNeuron, a developer tool meant to loosen Nvidia's control of PyTorch, so more AI workloads run on Amazon's own chips. Teams with long training pipelines should benchmark it before assuming Nvidia-only pricing.

Other

Platform updates rounded out the day: fine-tuning tooling from Together AI, Mercury 3 opening on a second routing service, a rebuilt agent layer inside Notion, and an easier path into AWS's managed model-building environment. Each one lowers the setup work between having a use case and running it.

• •Together Post-Training Together AI launched a post-training service for fine-tuning and alignment, claiming faster runs and lower costs than assembling a training stack yourself. It fits teams with domain data but no in-house training infrastructure.
• •Mercury 3 on OpenRouter Inception's Mercury 3 is routable on OpenRouter (a service that sends one request to many models) with a 1,000k context and pricing at $0.10 per million input tokens and $0.35 per million output. Cheap enough for fast agent loops.
• •Notion agents rebuilt Notion reworked its agents to do more work for less, tightening AI features inside its workspace. SMB owners can automate notes, summaries, and database updates without buying a separate tool.
• •SageMaker Studio in the new AWS console AWS now exposes SageMaker Studio, its managed notebook and model-building environment, through the new console experience, cutting setup steps for teams already running on AWS.

What this means for you

For Vibe Builders: You can stop paying premium rates for the boring parts: MiniMax M3.1 Flash is free on Vercel's AI Gateway until 20 October and Mercury 3 is on OpenRouter at $0.10 per million input tokens with a 1M-token context. Add a local vision model such as MiniCPM-V-4.7's gguf build and your agent can read screenshots without a cloud bill. Because the AI cloud host shakeout is expected, keep your model calls on a router so you can swap providers in one line.

For Non-techies: Two things changed for a small business today. Notion rebuilt its agents to do more work for less, so notes, summaries, and database updates can run inside the tool you already pay for, and Wabi launched as a free personal assistant you can download now. On the cost side, MiniMax's M3.1 Flash is free on Vercel's AI Gateway until 20 October, and its benchmark scores roughly doubled the value per dollar, which is the pressure that eventually reaches every tool you buy.

For Developers: Security is now the gating item: Microsoft is telling enterprises to treat employee-used agents as insider threats, so audit agent permissions and log outbound calls before you ship. On the platform side, Amazon's TorchNeuron tool targets Nvidia-only PyTorch pipelines, AWS put SageMaker Studio into its new console, and Together AI opened a post-training service for teams with domain data. Mercury 3's 1,000 tokens per second and 1M context make it worth benchmarking for latency-sensitive agent loops, and MiniMax M3.1 Flash hit 73.7 percent on DeepSWE v1.1 for under $0.30 per task.

What to watch next

Watch whether MiniMax keeps free pricing past 20 October, because that sets the floor rivals must answer. Track the neocloud shakeout too: if Microsoft and others keep abandoning chip deals, availability tightens for smaller providers. And watch Nvidia's Germany and European Broadcasting Union work for signs that industrial and media buyers become the next large customer base.

Sources

Hugging Face trending

• •Beam-1.0-501B-A23B by reflection trends on HuggingFace
• •Muse-Spark-Open-816B-A42B-NVFP4 by meta-models trends on HuggingFace
• •Timestral-209M by mistralai trends on HuggingFace
• •NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 by nvidia trends on HuggingFace
• •MiniCPM-V-4.7-35B-A3B-gguf by openbmb trends on HuggingFace

Vendor launches

• •Updates to Hobby and Pro Terms of Service
• •Germany Wants to Lead the World in Industrial AI: How NVIDIA Will Help
• •Mercury 3 now available on AI Gateway
• •NVIDIA Advances Agentic Security for Media, From European Public Service Broadcasting to Global Content Production
• •We’re supporting nature restoration through sustainable farming in Brazil.
• •MiniMax M3.1 Flash now available on AI Gateway for free

Industry news

• •The neocloud shakeout is coming, and Nvidia has the most to lose
• •People are using AI to make their smart TVs dumb again
• •Laplace Intelligence red-teams robots and self-driving cars before they hit the real world
• •“We were looking at it wrong”: The explosive claim that’s upended the AI world
• •Tolan snags $30M for its AI companion that helps users navigate life
• •Replika’s founder built the AI assistant she pitched at Disrupt 12 years ago
• •OpenAI Engineer Accused of Spying on Executives Ordered to Remain in Custody
• •MiniMax M3.1 Flash shows China's AI price race has no floor in sight
• •Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’
• •Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool

Other

• •Introducing Together Post-Training: Better models, faster runs, lower costs →
• •Inception: Mercury 3 now available on OpenRouter (1,000k context, $0.10/M in, $0.35/M out)
• •Agents, rebuilt to do more for less.
• •Developers can now access SageMaker Studio through the new AWS experience