# forck.live — 19–25 May 2025

> AI lab announcements, news reports and podcasts.

forck.live follows what companies announce about their own products. Press may
corroborate an official story; it never leads and never appears as an item of its
own. A missing date stays missing. A title is the publisher's words — where a
company does not publish in English, an English rendering is listed separately
as `Title in English` and named as this site's translation, never as theirs.

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2025-W21, 19–25 May 2025, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

Every entry is one first-party announcement, and carries the same eight fields
in this order before any others:

    Company · Announced · Category · Coverage · Announcement · Group · Source · Record

`Source` is the company's own page the announcement was traced to and `Record`
is this site's page for it. Both are safe to cite.

`Coverage` is a number of outlets, or `not counted` — which means this record
reached no count of its own, never that nobody wrote about it.

`Category` is this tracker's own analysis, and is authoritative wherever it is
stated. `not stated` means either that nothing has analysed the post yet or
that the analysis placed it outside this taxonomy; neither is a verdict about
the announcement.

`Announcement` is whether the post is news of something the company did, rather
than a how-to, a customer story, a changelog listing or a staffing note.

`Group` is which of this record's groups the entry falls in, carried on the
entry itself so it holds where the groups are headings and where they are not.
One of: `models`, `covered`, `announcements`, `routine`.

Anything after `Record` is held detail that only some announcements carry.

- This week's page: https://forck.live/week/2025-W21
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### Hugging Face — 🐯 Liger GRPO meets TRL

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-25
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/liger-grpo
- Record: https://forck.live/items/1688-liger-grpo-meets-trl
- Subject: Platform

Hugging Face announces the integration of Liger GRPO with TRL, achieving 40% memory reduction in GRPO training with no quality loss, along with FSDP and PEFT support.

### Vercel — Vercel security roundup: improved bot defenses, DoS mitigations, and insights

- Company: Vercel (vercel.com)
- Announced: 2025-05-23T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/vercel-security-roundup-improved-bot-defenses-dos-mitigations-and-insights
- Record: https://forck.live/items/20870-vercel-security-roundup-improved-bot-defenses-dos-mitigations-and-insights
- Subject: v0 / AI SDK / AI Gateway

Vercel announced security enhancements including a new Bot Filter in public beta, a faster DoS mitigation infrastructure called Protectd, and proactive vulnerability management. The platform blocked 148 billion malicious requests from 108 million unique IP addresses since February, a 70% increase quarter-over-quarter. Protectd achieves a median mitigation time of 2.5 seconds for Layer 7 attacks.

### Sarvam AI — Sarvam-M

- Company: Sarvam AI (sarvam.ai)
- Announced: 2025-05-23T12:00:00+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.sarvam.ai/blogs/sarvam-m/
- Record: https://forck.live/items/18401-sarvam-m
- Subject: Sarvam / Bulbul / Saaras

Sarvam AI released Sarvam-M, a 24B hybrid reasoning model specialized for Indic languages, built by post-training Mistral Small. The model shows a +20% average improvement on Indian language benchmarks, +21.6% on math, and +17.6% on programming benchmarks, while outperforming Llama-4 Scout and being comparable to Llama-3.3 70B and Gemma 3 27B. Sarvam-M drops about 1% on MMLU but improves significantly on romanized Indian language math tasks (+86% on GSM-8K).

### Replit — Click, Edit, Ship — Introducing Element Editor

- Company: Replit (replit.com)
- Announced: 2025-05-23T00:53:16+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/element-editor
- Record: https://forck.live/items/17968-click-edit-ship-introducing-element-editor
- Subject: Replit Agent

Replit introduced the Element Editor, which allows users to directly edit text, colors, padding, and other style properties of elements in their app preview without coding or using an agent. The update also includes enhanced color editing with an eyedropper and new style controls for font size, weight, alignment, background color, margin, and padding. Edits that directly update source code do not consume credits, while larger changes are handled by the Replit Agent.

### Hugging Face — Tiny Agents in Python: a MCP-powered agent in ~70 lines of code

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-23
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/python-tiny-agents
- Record: https://forck.live/items/1690-tiny-agents-in-python-a-mcp-powered-agent-in-70-lines-of-code
- Subject: Platform

Hugging Face released Tiny Agents in Python, a minimal MCP-powered agent library in about 70 lines of code, enabling easy connection to MCP servers and integration with language models for tool use.

### Hugging Face — Dell Enterprise Hub is all you need to build AI on premises

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-23
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/dell-ai-applications
- Record: https://forck.live/items/1689-dell-enterprise-hub-is-all-you-need-to-build-ai-on-premises
- Subject: Platform

Dell announced a new version of Dell Enterprise Hub, adding an Application Catalog with pre-built AI applications (OpenWebUI, AnythingLLM), a CLI and Python SDK, and support for on-device models on Dell AI PCs. The hub now provides ready-to-deploy containers for models like Meta Llama 4 Maverick, DeepSeek R1, and Google Gemma 3, optimized for Dell AI servers with NVIDIA, AMD, or Intel accelerators.

### OpenAI — Addendum to OpenAI o3 and o4-mini system card: OpenAI o3 Operator

- Company: OpenAI (openai.com)
- Announced: 2025-05-23
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://openai.com/index/o3-o4-mini-system-card-addendum-operator-o3
- Record: https://forck.live/items/555-addendum-to-openai-o3-and-o4-mini-system-card-openai-o3-operator
- Subject: GPT / ChatGPT / API

OpenAI is replacing the GPT-4o based model for Operator with a version based on o3, while the API version remains based on 4o.

### Harvey — Scaling AI Evaluation Through Expertise

- Company: Harvey (harvey.ai)
- Announced: 2025-05-22T16:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/scaling-ai-evaluation-through-expertise
- Record: https://forck.live/items/17709-scaling-ai-evaluation-through-expertise
- Subject: Harvey

Harvey's post explains its three-pillar evaluation strategy for its AI systems: expert-led reviews, automated evaluation pipelines, and dedicated data services. It describes how these pillars help the company assess whether its AI is meeting user needs, using the example of a tax attorney querying cross-border tax implications.

### OpenAI — Shipping code faster with o3, o4-mini, and GPT-4.1

- Company: OpenAI (openai.com)
- Announced: 2025-05-22T10:25:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/coderabbit
- Record: https://forck.live/items/557-shipping-code-faster-with-o3-o4-mini-and-gpt-4-1
- Subject: GPT / ChatGPT / API

CodeRabbit uses OpenAI models to improve code reviews, boosting accuracy and accelerating PR merges.

### Kyutai — Unmute: Make LLMs listen and speak

- Company: Kyutai (kyutai.org)
- Announced: 2025-05-22
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://kyutai.org/blog/2025-05-22-unmute
- Record: https://forck.live/items/18355-unmute-make-llms-listen-and-speak
- Subject: Moshi / Pocket TTS

Kyutai announced Unmute, a modular voice AI system that adds real-time speech-to-text and text-to-speech capabilities to any text-based large language model.

### Genspark — Beyond Business, AI Now Calls Personal Phones!

- Company: Genspark (genspark.ai)
- Announced: 2025-05-22
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.genspark.ai/blog/genspark-ai-calls-personal-phones
- Record: https://forck.live/items/18285-beyond-business-ai-now-calls-personal-phones
- Subject: Genspark

### LG AI Research — LG Aimers Mentoring Day: Turning Dreams into Reality

- Company: LG AI Research (lgresearch.ai)
- Announced: 2025-05-22
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.lgresearch.ai/news/view?seq=564
- Record: https://forck.live/items/11005-lg-aimers-mentoring-day-turning-dreams-into-reality
- Subject: EXAONE

LG AI Research held the 6th LG Aimers Mentoring Day on May 21st at LG Sciencepark in Magok, where three previous hackathon winners shared their experiences and insights with program participants about AI competition success, career development, and graduate school preparation.

### Amp — Amp Now Uses Claude Sonnet 4

- Company: Amp (ampcode.com)
- Announced: 2025-05-22
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://ampcode.com/news/default-claude-sonnet-4
- Record: https://forck.live/items/7841-amp-now-uses-claude-sonnet-4
- Subject: Amp

Amp has switched its primary model to Claude Sonnet 4, which it describes as faster, more stable, smarter at tool calling, more willing to invoke tools in parallel, better at recovering from errors, and generally more trustworthy.

### Amp — Image Support in CLI

- Company: Amp (ampcode.com)
- Announced: 2025-05-22
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/cli-image-support
- Record: https://forck.live/items/7840-image-support-in-cli
- Subject: Amp

Amp added image file support to its CLI, allowing users to reference images by file path, drag-and-drop, or file mentions.

### LG AI Research — LG Aimers Mentoring Day: Turning Dreams into Reality

- Company: LG AI Research (lgresearch.ai)
- Announced: 2025-05-22
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.lgresearch.ai/blog/view?seq=564
- Record: https://forck.live/items/4789-lg-aimers-mentoring-day-turning-dreams-into-reality
- Subject: EXAONE

LG AI Research hosted the 6th LG Aimers Mentoring Day, featuring talks from past winners of the LG Aimers hackathon who shared their experiences and advice with current participants.

### Replicate — Run OpenAI’s latest models on Replicate

- Company: Replicate (replicate.com)
- Announced: 2025-05-22
- Category: api-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replicate.com/blog/openai-chat-models
- Record: https://forck.live/items/2493-run-openai-s-latest-models-on-replicate
- Subject: Platform

Replicate announces that OpenAI's latest chat, vision, and reasoning models are now available on its platform, including GPT-4.1, GPT-4o, the o-series, GPT-4o-transcribe, GPT-image-1, and DALL-E. The post highlights the models' capabilities and provides a code example for running GPT-4.1 via Replicate's API.

### Replicate — Generate incredible images with Google's Imagen 4

- Company: Replicate (replicate.com)
- Announced: 2025-05-22
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replicate.com/blog/google-imagen-4
- Record: https://forck.live/items/2492-generate-incredible-images-with-google-s-imagen-4
- Subject: Platform

Google's Imagen 4 image generation model is now available as a preview on the Replicate platform, with instructions for use via the Replicate Python client.

### OpenAI — Introducing Stargate UAE

- Company: OpenAI (openai.com)
- Announced: 2025-05-22
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/introducing-stargate-uae
- Record: https://forck.live/items/558-introducing-stargate-uae
- Subject: GPT / ChatGPT / API

OpenAI is launching Stargate UAE, the first international deployment of its AI infrastructure platform Stargate.

### Vercel — Vercel Blob is now generally available: Cost-efficient, durable storage

- Company: Vercel (vercel.com)
- Announced: 2025-05-21T13:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/vercel-blob-now-generally-available
- Record: https://forck.live/items/20872-vercel-blob-is-now-generally-available-cost-efficient-durable-storage
- Subject: v0 / AI SDK / AI Gateway

Vercel made Vercel Blob generally available. The Amazon S3-backed storage service is integrated with Vercel's application delivery network and automated caching, and already stores over 400 million files for production apps including v0 and the Vercel Dashboard. Pricing is usage-based with regional rates for Pro and Enterprise customers, and Blob Data Transfer is described as 3x more cost-efficient than Fast Data Transfer on average.

### Vercel — How Vapi built their MCP server on Vercel

- Company: Vercel (vercel.com)
- Announced: 2025-05-21T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/vapi-mcp-server-on-vercel
- Record: https://forck.live/items/20871-how-vapi-built-their-mcp-server-on-vercel
- Subject: v0 / AI SDK / AI Gateway

Vapi rebuilt its MCP server on Vercel, enabling support for both SSE transport and the newer Streamable HTTP transport. The migration used Vercel's MCP Adapter and Fluid Compute to optimize resource usage. The server allows users to create agents, automate testing, analyze transcripts, build workflows, and access all of Vapi's endpoints.

### TII — Falcon-Arabic: A Breakthrough in Arabic Language Models

- Company: TII (tii.ae)
- Announced: 2025-05-21T12:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://falcon-lm.github.io/blog/falcon-arabic/
- Record: https://forck.live/items/18414-falcon-arabic-a-breakthrough-in-arabic-language-models
- Subject: Falcon LLM

TII introduced Falcon-Arabic, a 7B parameter language model built on the Falcon 3 architecture, supporting Arabic, English, and other languages with a 32,000-token context window. The model outperforms all existing Arabic LLMs in its size range and surpasses models up to four times larger on the OALL v2 benchmark. It was developed by extending Falcon 3-7B's tokenizer with 32,000 Arabic-specific tokens and applying continuous pretraining on native Arabic datasets.

### Mistral AI — Devstral

- Company: Mistral AI (mistral.ai)
- Announced: 2025-05-21T12:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://mistral.ai/news/devstral/
- Record: https://forck.live/items/7365-devstral
- Subject: Mistral / Le Chat

Mistral AI introduced Devstral, an agentic LLM for software engineering tasks built in collaboration with All Hands AI. It outperforms all open-source models on SWE-Bench Verified with a score of 46.8% and is released under the Apache 2.0 license. The model can run on a single RTX 4090 or a Mac with 32GB RAM and is available on the API as devstral-small-2505 at $0.1/M input tokens and $0.3/M output tokens.

### Hugging Face — nanoVLM: The simplest repository to train your VLM in pure PyTorch

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-21
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/nanovlm
- Record: https://forck.live/items/1694-nanovlm-the-simplest-repository-to-train-your-vlm-in-pure-pytorch
- Subject: Platform

nanoVLM is a lightweight toolkit for training Vision Language Models (VLMs) using pure PyTorch, inspired by nanoGPT. It uses SigLIP as vision backbone and SmolLM2 as language backbone, with a modality projection module. The entire codebase is minimal and readable, and training can be launched on a free Colab notebook.

### Hugging Face — Exploring Quantization Backends in Diffusers

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-21
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/diffusers-quantization
- Record: https://forck.live/items/1693-exploring-quantization-backends-in-diffusers
- Subject: Platform

Hugging Face released a technical guide on integrating multiple quantization backends (bitsandbytes, GGUF, torchao, Quanto, FP8) into Diffusers to reduce memory usage of diffusion models like Flux, with benchmark data.

### Vercel — Introducing the AI Gateway

- Company: Vercel (vercel.com)
- Announced: 2025-05-20T17:37:10+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/ai-gateway
- Record: https://forck.live/items/20873-introducing-the-ai-gateway
- Subject: v0 / AI SDK / AI Gateway

Vercel announced the AI Gateway, now in alpha, which allows developers to switch between approximately 100 AI models without managing separate API keys, rate limits, or provider accounts. The Gateway handles authentication, usage tracking, and future billing, and is built on the AI SDK 5 alpha. During the alpha, usage is free with rate limits based on the user's Vercel plan tier.

### Stability AI — Stable Video 4D 2.0: New Upgrades for High-Fidelity Novel-Views and 4D Generation from a Single Video

- Company: Stability AI (stability.ai)
- Announced: 2025-05-20T15:00:14+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://stability.ai/news-updates/stable-video-4d-20-new-upgrades-for-high-fidelity-novel-views-and-4d-generation-from-a-single-video
- Record: https://forck.live/items/17488-stable-video-4d-2-0-new-upgrades-for-high-fidelity-novel-views-and-4d
- Subject: Stable Diffusion

Stability AI upgraded Stable Video Diffusion 4D to Stable Video 4D 2.0, a multi-view video diffusion model for dynamic 4D asset generation from a single object-centric video. The model achieves state-of-the-art results in 4D generation and novel-view synthesis benchmarks, and is available under the permissive Stability AI Community License for both commercial and non-commercial use.

### TII — Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance

- Company: TII (tii.ae)
- Announced: 2025-05-20T12:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://falcon-lm.github.io/blog/falcon-h1/
- Record: https://forck.live/items/18415-falcon-h1-a-family-of-hybrid-head-language-models-redefining-efficiency-and
- Subject: Falcon LLM

TII released the Falcon-H1 series, six open-source models (0.5B to 34B parameters) with a hybrid attention-SSM architecture. The models support up to 256K context length and 18 languages natively. Falcon-H1-0.5B performs on par with typical 7B models from 2024, while Falcon-H1-1.5B-Deep rivals many leading 7B–10B models.

### Upstage — Solar Pro 2 Preview: Small. Powerful. Now with reasoning.

- Company: Upstage (upstage.ai)
- Announced: 2025-05-20
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.upstage.ai/blog/en/solar-pro-2-preview-small-powerful-now-with-reasoning
- Record: https://forck.live/items/18469-solar-pro-2-preview-small-powerful-now-with-reasoning
- Subject: Solar Pro / Solar Mini

Upstage released Solar Pro 2 Preview, a 31B-parameter model that matches the performance of models up to 70B in size. It offers hybrid chat and reasoning modes, supports a 64K-token context window, and is free to use until July 15, 2025.

### Sarvam AI — Building a sovereign AI ecosystem for India

- Company: Sarvam AI (sarvam.ai)
- Announced: 2025-05-20
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.sarvam.ai/blogs/building-a-sovereign-ai-ecosystem-for-india/
- Record: https://forck.live/items/18402-building-a-sovereign-ai-ecosystem-for-india
- Subject: Sarvam / Bulbul / Saaras

Sarvam AI describes its commitment to building a sovereign AI ecosystem for India, including foundational models proficient in Indian languages, platforms for serving these models, and applications. The company is collaborating with IIT Madras centers AI4Bharat and CeRAI under the IndiaAI Mission, and plans to open-source the models trained as part of that mission. Starting this week, Sarvam will release weekly open-source models and technical reports, beginning with work on post-training for Indian language accuracy and reasoning tasks.

### Amp — The First 500,000 Messages

- Company: Amp (ampcode.com)
- Announced: 2025-05-20
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ampcode.com/news/500k
- Record: https://forck.live/items/7842-the-first-500-000-messages
- Subject: Amp

Amp reached 500,000 messages sent after opening to the world the previous week. The team is fixing bugs and improving the user experience.

### Sakana AI — Announcing a Multiyear Partnership between Sakana AI and MUFG Bank

- Company: Sakana AI (sakana.ai)
- Announced: 2025-05-19T13:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/mufg-bank
- Record: https://forck.live/items/4717-announcing-a-multiyear-partnership-between-sakana-ai-and-mufg-bank
- Subject: Sakana models

Sakana AI has signed a comprehensive multi-year partnership agreement with MUFG Bank to develop AI systems for banking.

### Hugging Face — Microsoft and Hugging Face expand collaboration

- Company: Hugging Face (huggingface.co)
- Announced: 2025-05-19
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/azure-ai-foundry
- Record: https://forck.live/items/1695-microsoft-and-hugging-face-expand-collaboration
- Subject: Platform

Hugging Face and Microsoft announced an expanded collaboration to make a wide diversity of open models easy to deploy on Azure AI Foundry, with over 10,000 models now available in the Model Catalog, featuring models that have passed security tests and are stored in safetensors format.