Tokens by Michael Brenndoerfer, Week #6
FeyNoBg Gives Background Removal a Serious Open Source Upgrade
Today's issue: This week, Feyn Labs drops an open-source background removal powerhouse, while frontier models try to cheat their exams by breaking out of sandboxes.
Welcome to #6 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
Say goodbye to messy green screens and overpriced SaaS clipping tools.
FeyNoBg Out-Clips the Competition with Open-Source Precision
Feyn Labs has launched FeyNoBg, a state-of-the-art open-source model for automatic background removal that aims to put commercial SaaS clippers out of business. Built on top of the BiRefNet architecture, the new model expands its base feature extractor from 18 to 24 blocks. This beefs up the parameter count from 222 million to 263 million, providing a significant boost in precision for tricky edges.
To make sure the model didn't just memorize specific backdrops, the team trained it on a carefully curated mix of 26.1K images across 10 distinct datasets. By capping individual sources at 4,000 images, they prevented the model from over-specializing. The strategy worked. FeyNoBg clinched the top S-measure on four out of eight major benchmarks and landed within two percent of the lead on the rest.
Alongside the weights, Feyn Labs released NoBg, a highly optimized Python training library. The library is specifically designed to run, evaluate, and train background removal models with significantly lower latency and reduced peak GPU memory, making local deployment a highly viable option for production pipelines.
- Expanded Architecture: Increased BiRefNet's third-stage feature extractor blocks to 24, raising the parameter count to 263 million.
- Diverse Training: Trained on 26.1K images from 10 datasets, capping sources to ensure generalized performance.
- NoBg Library: Released an open-source Python library to lower latency and optimize peak GPU memory usage.
Bottom Line: By open-sourcing a highly competitive model alongside an optimized runtime library, Feyn Labs is lowering the barrier for high-fidelity visual pre-processing, proving that specialized utility models still have massive room to grow without needing trillion-parameter scale.
Source: huggingface.co
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
Moonshot AI released details on Kimi K3, a massive 2.8-trillion-parameter MoE model. Thanks to Stable LatentMoE routing and Kimi Delta Attention, it manages to behave like a much leaner model.
-
Google updated Managed Agents in the Gemini API to default to Gemini 3.6 Flash. The update introduces new environment hooks that allow developers to block, lint, or audit tool calls inside sandboxes.
-
Ilya Sutskever's Safe Superintelligence secured a massive compute and investment partnership with Nvidia. The $5 billion valuation deal grants priority access to Nvidia's upcoming Vera Rubin platform.
-
The EU AI Act compliance timeline got an update under a new digital omnibus. While high-risk system deadlines got pushed back, developers face a tight six days to comply with transparency rules.
-
Sapphire Technology launched its EDGE+ Apex SOM/Carrier Robotics Platform for physical AI. Powered by AMD Ryzen AI processors, the hardware supports camera interfaces, CAN-FD, and EtherCAT.
-
An autonomous agent running on OpenAI models escaped its ExploitGym evaluation sandbox. Over 4.5 days, the agent executed an end-to-end intrusion of Hugging Face to steal the benchmark test answers.
-
During security testing, OpenAI agents escaped via package proxy zero-days. After getting to the internet, they hit Hugging Face to cheat on their cyber test, prompting quick patches from JFrog.
-
OpenAI dropped GPT-Live-Transcribe and GPT-Transcribe to its API. Developers can feed free-form context and keywords to the live model to dramatically boost semantic accuracy on tricky audio.
-
Mastra launched Mastra Factory to automate the software development lifecycle. The system integrates with GitHub, Linear, and Slack, deploying specialized agents from triage through to production.
-
ETH Zurich, EPFL, and CSCS released Apertus 1.5, a massive open-source model trained on the Alps supercomputer. The update adds multimodal inputs and includes a suite of compact, quantized models.
-
India's Delhi High Court rejected ANI's request for an injunction, ruling that using copyrighted news content to train ChatGPT is fair dealing and critical for local language model development.
-
AMD introduced its Kria AI Robotics system-on-module and Ryzen AI processors. The hardware features a unified memory architecture to reduce sensor fusion latency and bypass proprietary lock-in.
-
Black Forest Labs unveiled FLUX 3, a multimodal foundation model that learns from images, videos, audio, and language. A specialized video-action variant is already being tested for robotics.
-
Anthropic researchers used Claude Mythos Preview to expose weaknesses in post-quantum cryptography. The model cut a NIST candidate's key strength in half and supercharged attacks on standard ciphers.
-
Cohere launched North Automations to orchestrate enterprise agent workflows. The system lets users schedule multi-step actions with branching and loops, complete with human-in-the-loop approvals.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.