DeepSeek put frontier-class open weights on Hugging Face with no announcement — the same week ChatGPT's image model melted OpenAI's GPUs

WEEK OF MAR 22 – 28, 2025

DeepSeek put frontier-class open weights on Hugging Face with no announcement — the same week ChatGPT's image model melted OpenAI's GPUs

V3-0324 appeared under MIT with no blog post and immediately became the strongest open-weight model on the board. Native image generation pulled about a million users in an hour, and OpenAI adopted the open standard a rival wrote.

The split in this week is almost too neat. On one side, DeepSeek quietly uploaded a 671-billion-parameter model to Hugging Face under MIT — no announcement, no technical report, no changelog — and independent evaluators found it had become the best open-weight model available, beating every closed non-reasoning model on a composite quality score. On the other, Sam Altman was asking users to chill on generating images because OpenAI's GPUs were melting.

In between, OpenAI adopted the Model Context Protocol, an open standard written by its competitor Anthropic, across the Agents SDK and promised it for the ChatGPT desktop app and Responses API. Three labs, three entirely different theories of where the moat lives — and this week they were all on display.

WED · Mar 26, 2025mcpopenaianthropicagentsstandardsinteroperability

OpenAI adopts MCP, the open agent standard written by its rival Anthropic

Sam Altman announced that OpenAI would add support for Anthropic's Model Context Protocol across its products, including the ChatGPT desktop app, calling it something people love. Support landed in the Agents SDK the same day, with the ChatGPT desktop app and Responses API to follow. Anthropic's chief product officer Mike Krieger welcomed it publicly, describing MCP as a thriving open standard with thousands of integrations.

MCP lets developers expose data sources as servers and connect apps and workflows as clients, giving models two-way access to the systems where data actually lives. Since Anthropic open-sourced it, Block, Apollo, Replit, Codeium and Sourcegraph had already added support.

OpenAI adopts MCP, the open agent standard written by its rival Anthropic — TechCrunch

WHY IT MATTERS

A genuinely rare event: the two leading US labs agreeing on shared plumbing for tool use. For anyone building agents, this removes the worst kind of bet — which lab's tool protocol to write against — and it means an MCP server you build is now usable by Claude, ChatGPT and a growing list of developer tools rather than being locked to one vendor. It also quietly sets the precedent that agent interoperability is being settled by open standards rather than by whoever has the largest install base, which is worth remembering the next time someone proposes a proprietary tool-calling format.

TechCrunch — OpenAI adopts rival Anthropic's standard for connecting AI models to data

TUE · Mar 25, 2025openaiimage generationgpt-4ocomputecopyrightscale

GPT-4o generates images natively and the internet breaks — about a million users in an hour, then rate limits

OpenAI added native image generation to GPT-4o, replacing DALL-E 3 as ChatGPT's default image generator and reclassifying DALL-E as legacy. Rather than handing requests to a separate diffusion service, image output is produced by the same transformer that handles text and vision, which is why the model can follow detailed instructions, render legible text inside images, edit an existing picture, and hold visual consistency across a conversation in a way earlier tools could not.

The reaction was the story. Altman said roughly a million new users arrived in about an hour, against about five days for ChatGPT's first million in 2022 — 'biblical demand' of a kind he said he had never seen — and then asked users to ease off because the team needed sleep and the GPUs were melting. OpenAI temporarily delayed the free-tier rollout, throttled generation, and moved free users toward three generations per day. Usage roughly doubled in the weeks around the launch; Altman cited around 500 million weekly active users by April, against 400 million in February.

What went viral was a style: users redrew family photos and memes in the visual language of Studio Ghibli, which ignited a copyright and attribution argument within days. OpenAI added a refusal that triggers when a user attempts to generate an image in the style of a living artist, and critics resurfaced a 2016 clip of Hayao Miyazaki calling an AI-driven animation demo an insult to life itself. The feature also had a documented bug — the model struggled to render the word 'sexy women' at all, which Altman called a bug to be fixed.

GPT-4o generates images natively and the internet breaks — about a million users in an hour, then rate limits — i.insider.com

WHY IT MATTERS

Two lessons worth keeping. The architecture one: image generation inside the main multimodal model is what makes conversational editing work, and it is the same direction Google took with Gemini 2.0 Flash two weeks earlier — native modality beats service handoff. The operational one: consumer generative AI demand can outrun provisioned capacity within hours, and the only response is rate limiting. Anyone designing a product on top of a shared image or video endpoint should assume their ceiling is set by someone else's GPU supply, and that a viral moment is a capacity incident rather than a growth event.

Business Insider — what you need to know about OpenAI, Google and Anthropic's latest moves OpenAI adds native image generation to GPT-4o

TUE · Mar 25, 2025googlegemini 2.5reasoninglong contextbenchmarkscoding

Google releases Gemini 2.5 Pro, a thinking model with a million-token context

Google put Gemini 2.5 Pro out as an experimental preview under the identifier gemini-2.5-pro-exp-03-25 and made it available to all Gemini users. It is a thinking model that reasons before responding, covers text, audio, images, video and code, and carries a 1M-token context window with a 64K-token output limit.

It arrived at the top of the leaderboards and stayed there for months. On the June 2025 model card, Gemini 2.5 Pro GA posts 88.0% on AIME 2025, 86.4% on GPQA Diamond, 82.2% on Aider Polyglot, 54.0% on SimpleQA and 82.0% on MMMU. The experimental 03-25 build alone scored 86.7% on AIME and 84.0% on GPQA Diamond at launch. Google said all new Gemini models would include reasoning by default.

WHY IT MATTERS

The month's second half was a straight race at the top of the reasoning tier, and Gemini 2.5 Pro made the strongest single showing — enough that developers described it as the best model for code within days, and the 1M context was real rather than nominal. Combined with Gemma 3 two weeks earlier, Google shipped both the strongest open model and the strongest closed model in the same month, which is the clearest illustration of the two-track strategy every major lab now runs: give away the weights that set your floor, sell the API that sets your ceiling.

Gemini 2.5 Pro model card (PDF, Google DeepMind)AI Wiki — Gemini 2.5 Pro

MON · Mar 24, 2025open weightsdeepseekmitmoereasoningcostself-hosting

DeepSeek silently uploads V3-0324 under MIT and takes the open-weights crown

A file called DeepSeek-V3-0324 appeared on Hugging Face under DeepSeek's account with an MIT license and no accompanying announcement, technical report or changelog. Community members and benchmarking services found it within hours. It is a 671-billion-parameter mixture-of-experts model activating about 37 billion parameters per forward pass, keeping V3's multi-head latent attention, auxiliary-loss-free load balancing and multi-token prediction objective, with a 128K context window and a July 2024 knowledge cutoff.

The gains came from a better post-training pipeline rather than a new architecture, and they are large. MMLU-Pro rose from 75.9 to 81.2, GPQA from 59.1 to 68.4, AIME from 39.6 to 59, and LiveCodeBench from 39.2 to 49.2, with WebDev Arena improving by about 240 points and noticeably better front-end code and function calling. The license changed too: the original V3 shipped under DeepSeek's own license, while this checkpoint moved to MIT, the same terms as R1. Artificial Analysis flagged it as the first open-weight model to exceed all commercially available non-reasoning proprietary models on their composite quality score.

The release format is itself the statement. By uploading weights with no fanfare, DeepSeek bypassed the press cycle that Western labs build major launches around — and it did so weeks into a US debate in which OpenAI had accused DeepSeek to Congress of free-riding on the capabilities developed by American frontier labs.

WHY IT MATTERS

This is a headline release and should be read as one. A 671B MoE under MIT — permissive for commercial use, fine-tuning, redistribution and self-hosting — that outranks every closed non-reasoning model on an independent composite score is the clearest possible demonstration that open weights are not a tier behind. The cost floor consequence is the part that matters commercially: when frontier-adjacent capability is downloadable under MIT, the price any lab can charge for the non-reasoning tier is bounded by the cost of running it yourself. And the timing is the detail worth sitting with — the same quarter US labs were briefing Congress about distillation and free-riding, DeepSeek simply handed the weights over, which is a harder competitive move to argue against than any benchmark.

Hugging Face — deepseek-ai/DeepSeek-V3-0324 (MIT)HPCwire — DeepSeek-V3-0324 quietly lands on Hugging Face AI/TLDR — DeepSeek-V3-0324 specs and benchmarks