Easy AI Daily News Roundup | December 17, 2025

Model Releases and Updates

• Xiaomi MiMo-V2-Flash: 309B-parameter MoE model with 15B active parameters, supporting 150 tokens/s inference and 256K long context. Achieves SOTA on SWE-Bench. Currently free on OpenRouter.
• Links: Model details | OpenRouter page
• OpenAI GPT Image 1.5: New core image model for ChatGPT Images with more precise editing and faster generation. Tops multiple benchmarks; available to all ChatGPT users and via API.
• Links: Announcement | API docs
• NVIDIA Nemotron-Cascade: 8B/14B models using Cascade RL, reaching 43.1% pass@1 on SWE-Bench Verified. Open-sourced.
• Link: Model details
• Meta SAM Audio (open source): Separates specific sounds from complex audio using text, visual, or time-span prompts; includes models, benchmarks, and a paper.
• Links: HF collection | Announcement
• AllenAI Molmo 2: Video multimodal model built on SigLIP2 and Qwen3; leads open models on video pointing/counting tasks. Apache-2.0 license.
• Link: Announcement
• Apple SHARP: Generates 3D Gaussian models from a single image in ~1 second — about 1000x faster than diffusion models with better perceptual fidelity.
• Link: Paper summary
• Claude Code 2.0.70 (Anthropic): 13 CLI changes, 3x memory usage improvement, input-clearing fixes.
• Link: Changelog
• Manus.im 1.6: Rolled out to all users; platform improvements (details undisclosed).
• Link: Announcement

Benchmarks and Performance

• MiMo-V2-Flash SOTA on SWE-Bench: 73.4% on SWE-Bench Verified; 71.7% on multilingual tasks. Details
• GPT Image 1.5 tops image benchmarks: LMArena (1277), Design Arena (1344), AA Arena (1272). LMArena thread | Design Arena
• Google FACTS leaderboard launched: Gemini 3 Pro leads at 68.8%; multimodal tasks remain challenging (~47%). Thread
• OpenAI FrontierScience: New open benchmark for PhD-level scientific reasoning across physics, chemistry, and biology. Announcement

Open Source and Ecosystem

• Nemotron 3 Nano now available on Ollama and MLX/LM Studio for local deployment. Ollama | LM Studio
• MiMo-V2-Flash free on OpenRouter: Link
• Mistral Small Creative launches experimentally on OpenRouter at $0.10/$0.30 pricing. Model page
• Unsloth AI testing of Nemotron 3 Nano finds it better than Qwen3 30B with lower failure rates and higher speed. HF model
• Eleuther community Synthema: A metalinguistic project for meaning compression, converting concepts into shorter symbolic grammar.

Multimodal and Audio

• MiniMax VTP visual tokenizer (open source): Improves diffusion model generation quality without extra compute. Announcement
• Runway Gen-4.5 now available to all paid users. Announcement

Tools and Infrastructure

• vLLM KV-aware load balancer: Rust-based, with consistent hashing, retries, and Kubernetes discovery. Announcement
• OpenHands production-grade Agent SDK: Supports tool calling and reasoning. Announcement
• Cline migrates to Vercel AI Gateway: Lower error rates; 10–40% better P99 latency. Announcement
• tinygrad strict AI PR policy: AI-generated PRs from unknown contributors will be closed immediately; contributors must understand every line of code.

Community and Discussion

• LMArena YouTube channel launched, with model explainers and free-model guides.
• LMArena December contest: "Holiday Celebration" theme, submissions close December 30, Discord Nitro prizes.
• Terence Tao on AGI: Current AI is "artificial general intelligence" in name only — relying on random or brute-force methods, it cannot reach human-level intelligence. Source
• MI6 chief on tech giants: Warns their influence rivals governments, calling for urgent regulation to counter disinformation risks. Source
• User quits nicotine and gaming addiction with ChatGPT: A Reddit user reports quitting within a week of conversations. Post

Discord Server Summaries

• BASI Jailbreaking: GPT-5 mini system messages, jailbreak prompts, a full DeepSeek jailbreak, and Simulation Override challenges.
• LMArena: GPT Image 1.5 version differences, Nano Banana Pro performance dips, Gemini 3 Flash latency, and model censorship debates.
• Unsloth AI: GRPO vs DPO, Nemotron vs Qwen comparisons, Colab H100 GPUs, and SAM Audio licensing questions.
• Cursor Community: HTTP 401 errors, token billing discrepancies, Agent window defaults, and a debugging-decay paper.
• OpenAI: Branched chats on mobile, FrontierScience, GPT Image 1.5, and GPT-5.2 toxicity concerns.

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