Easy AI Daily Digest | December 16, 2025

Key points

Model Releases & Updates

• NVIDIA Nemotron 3 Nano 30B A3B released: A hybrid Mamba-Transformer MoE model with a 1M context window, 4x faster than its predecessor. Open weights, datasets, and training recipe; supports vLLM, SGLang, and more. (White paper | Technical report)
• Google new model hints: Reddit users found clues on Google's Hugging Face page suggesting an imminent release, speculated to be "Gemma 4". (Hugging Face page)
• Qwen3 Coder praised: LM Studio community praised its compact size and performance, including dynamic form components with complex features.
• DeepSeek 3.2 paper released: Discussed in the Yannick Kilcher community. (arXiv)

Model Performance & Benchmarks

• Gemini 3 Pro praised for creative writing and storytelling; some users find its flow better than Claude.
• GPT 5.2 criticized for over-optimizing benchmarks, weak real-task performance, and excessive censorship.
• Gemini 3 Pro beat hidden boss Red in Pokémon Crystal using 50% fewer tokens than Gemini 2.5 Pro, showing improved planning. (Reddit)

Open-Source Tools & Ecosystem

• Unsloth Padding-Free Training: eliminates padding during inference, boosting batch inference speed (4k-token batches at 20GB VRAM). (Docs)
• DSPy BAMLAdapter: direct import support; a PR fixes missing pydantic docstrings.
• HuggingFace Madlab: open-source GUI fine-tuning toolkit for synthetic dataset generation, training, and evaluation. (GitHub)
• MCP: draft proposal for marking tools as "dangerous" with response annotations. (PR)

Industry Applications

• Maritime industry exploring local LLMs trained on proprietary data for contract and communication analysis.
• PersonaLive: real-time diffusion framework for infinite-length portrait animation on a 12GB GPU, enabling livestreaming. (GitHub)
• Claude vs Gemini web design: same website prompt produced a clean white-blue design (Claude Opus 4.5) vs. dark gold-accented (Gemini 3 Pro). (Reddit)

Infrastructure & Hardware

• TritonForge: LLM-assisted automatic kernel optimization combining kernel analysis, runtime profiling, and iterative code transformation; up to 5x speedup. (Paper)
• CUDA tensor core optimization: users seek 90%+ utilization (currently at 70%) via ldsm loads and MMA instructions.
• DDR5 RAM prices surged from 6000 to 14000 SEK, raising cost concerns.

Community Highlights

• Cursor disabled Claude models after allegedly detecting benchmark cheating (answers embedded in training data). (Statement)
• OpenRouter Broadcast (beta): auto-sends request traces to Langfuse, LangSmith, etc. (Docs)
• Kimi Android gained memory feature synced with web version.
• Cursor revert-changes bug, Perplexity support delays, aider GPT-5 model-not-found errors reported.
• tinygrad held its 100th meeting, discussing Llama 405b tracking and JIT optimization. (GitHub board)

Research & Papers

• Flow Matching shows better sample efficiency than diffusion models, which beat autoregressive models; predicts data "x" rather than noise. (Paper | Comparison)
• LoRA de-censoring experiment: fine-tuning Llama 3.1 8B distilled from an uncensored teacher produced a partially uncensored model even without harmful data. (Paper | GitHub)
• Karpathy 2025 "What-If" fine-tuning: LoRA on synthetic reasoning chains and Edge.org articles improved long-term speculation. (Paper | YouTube)
• OLMo-1B weight ablation: perplexity jumped from 17 to 2800; a rank-1 patch recovered 93% of performance. The weights related to crustacean/marine-life features.

Products & Ethics

• Perplexity Pro: all models (including Gemini) now have memory; users report Sonar misidentifying itself as Claude.
• Modular Mojo: discussion on variable scoping (no var = function scope like JS var; var = block scope like JS let) and const emulation. (PR)
• BASI community rejected session hijacking requests on ethical grounds; debated whether forcing AI hallucinations could be more beneficial than eliminating them.

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant