Digest

This week in open AI: on-device inference meets agentic breakthroughs

Sparse experts land on smartphones, web agents learn from live sites, and uncensored fine-tunes proliferate—while Anthropic files to go public and Nvidia preps 120B-parameter laptops.

ByAlex Sokoloff·June 2, 2026

Two trends collided this week: pushing frontier-scale models onto consumer hardware and teaching agents to navigate messier, real-world environments. Nvidia and Liquid AI both announced architectures that fit double-digit-billion-parameter models into phone and laptop memory, while researchers demonstrated that reinforcement learning on live websites and fairness-aware policy synthesis can unlock new agent capabilities. Meanwhile, the open fine-tune ecosystem continued its rapid expansion.

In this digest

OpenWebRL-4B reaches 67% success on live-web tasks with minimal supervised data — Training web agents on real sites with only 400 seed examples just hit 67% success—online RL may finally beat the sim-to-real gap.

MobileMoE scales sparse experts to smartphones, cuts inference by 2–4× — Sparse experts cut smartphone inference time in half without ballooning memory—MoE design is now a mobile-first architecture question.

Autoresearch agent discovers fairness mechanisms emerge only under maximin objectives — An autonomous researcher agent just proved fairness mechanisms only emerge when you optimize for equity, not efficiency—a result no human study isolated.

Liquid AI's LFM2.5-8B-A1B runs on phones with 1.5B active parameters — Liquid AI's 8B-parameter MoE activates only 1.5B per pass and ships for phones, laptops, and robots—on-device just got a new baseline.

ProRL corrects gradient bias in recommendation path planning — ProRL fixes two gradient-bias bugs that have plagued proactive recommenders, beating baselines on three real datasets—RL debugging pays off.

Claude Opus 4.8 catches 4× more code bugs, leads on browser-agent benchmarks — Claude Opus 4.8 catches four times more code bugs and leads browser-agent benchmarks—Anthropic is betting on reliability over raw scale.

Nvidia RTX Spark Superchip brings 120B local LLMs to Windows PCs — Nvidia's RTX Spark Superchip pairs Arm cores with 128 GB unified RAM to run 120B models in FP4 on Windows PCs—local inference just leapfrogged a generation.

Illustrious XL 1.0 Improved Uncensored v3.0 SDXL fine-tune lands on HuggingFace — Illustrious XL 1.0 Improved Uncensored v3.0 lands on HuggingFace with safetensors—anime-focused SDXL fine-tunes remain a prolific niche.

Also this week

Anthropic files confidential S-1 with SEC, moving toward IPO — Anthropic filed a confidential S-1 with the SEC, the first formal step toward a public offering; timing and valuation remain undisclosed.
BitsMoE cuts Qwen3-30B 2-bit quantization time by 12× with spectral bit allocation — BitsMoE uses SVD-based spectral bit allocation to quantize Qwen3-30B to 2 bits 12× faster, gaining 27.83 points over GPTQ.
Nova Anime AM hits 8,000 downloads with optimized low-step sampling — Nova Anime AM, an Anima checkpoint optimized for low-step anime generation, hit 8,060 downloads on Civitai.
Google used Gemini to produce I/O 2026 conference — Google disclosed it used Gemini to handle production workflows for I/O 2026, the first public confirmation of AI staging the flagship conference.
Qwen-Image NSFW LoRA launches on HuggingFace with Diffusers support — Gigo1990 released an uncensored LoRA for Alibaba's Qwen-Image on HuggingFace, enabling local text-to-image without safety filters.
osmKeye-VL-2.0-30B abliterated weights land in three quantizations on HuggingFace — osmapi published abliterated weights for osmKeye-VL-2.0-30B-A3B in full precision and two MLX quantizations on HuggingFace.