Easy AI Daily | January 29, 2026

Models & Capabilities

Kimi K2.5: A New Open-Model Leader

Moonshot's Kimi K2.5 ranks #1 among open models on the LMArena/Arena text leaderboard, with standout STEM and coding performance. Community testing puts its coding near Claude Opus 4.5 quality at roughly 10–20% of the price, and it can run locally (1T MoE; ~240GB GGUF after Unsloth quantization). Known drawbacks: verbose wording, high token usage, and occasional hallucinations.

Links: Text Arena leaderboard | Kimi announcement | Kimi-K2.5 GGUF (Unsloth) | Local run guide (Reddit)

Kimi K2.5 System Prompt & Toolchain Leaked

The complete system prompt and tool configuration (~5K tokens) of Kimi K2.5 were published, including memory CRUD protocols, tool schemas, context injection, and safety rules. It's a valuable reference for open-source projects aiming to replicate high-quality assistant personas, memory mechanisms, and long-conversation stability.

Links: Leaked repo | Reddit discussion

Trinity Large: 400B Open MoE

Arcee / Prime Intellect / Datology released Trinity Large: a 400B sparse MoE activating only 13B parameters at inference — 256 experts, 4 per token (~1.56% routing). Open weights under a permissive license, already integrated in Cline and OpenRouter.

Links: Tech report | OpenRouter preview

DeepSeek-OCR 2 Released

DeepSeek-OCR 2 introduces DeepEncoder V2 with a visual causal flow that "understands the whole page first, then decides reading order", significantly reducing edit distance. Community hopes V2 fixes V1's misses on headers/footers and light-colored text; users are seeking a hosted API.

Links: Announcement | Online demo

Z-Image Base / Turbo: New Open Text-to-Image Models

Alibaba Tongyi released Z-Image Base (more steps, finer quality) and Turbo (~7s generation, more photorealistic). Community reports it runs on 12GB VRAM; the Stable Diffusion community calls it "like a new-generation SDXL".

Links: Model page | Base vs Turbo discussion

DeepMind Open-Sources AlphaGenome Weights

Google DeepMind opened the AlphaGenome model and weights for predicting the molecular effects of genetic variants, citing over a million daily API calls and 3,000+ internal users. Weights are on a Hugging Face collection.

Links: Announcement | Weights links

NVIDIA Releases NVFP4 Nemotron 3 Nano

NVIDIA's NVFP4-precision Nemotron 3 Nano claims up to 4x BF16 throughput on Blackwell B200 with 99.4% precision retained via quantization-aware distillation (QAD). vLLM already supports the format.

Links: NVIDIA tweet | vLLM NVFP4 support

MongoDB LEAF: Distilling Embeddings to CPU-Scale

MongoDB Research's LEAF computes embeddings offline with a large model and serves queries with a small aligned model, retaining ~96% quality while shrinking model size 5–15x and boosting query throughput up to 24x on CPU/edge devices.

Frontier Model "Personality" Division of Labor

An analysis splits frontier models into exploration vs. exploitation: GPT-5.2 suits broad search and deep reasoning; Claude Opus 4.5 is stable and token-efficient — research workflows favor GPT, production workloads favor Claude.

Agents & Tooling

• Agent Skills ecosystem forming: DeepLearning.AI + Anthropic launched an "Agent Skills" course; LangChain is pushing Skills; Hugging Face's upskill case shows strong-model traces can lift weak-model accuracy up to 45% on some CUDA kernel tasks — but results vary and must be evaluated per model.
• Agent evaluation is shifting to trajectories: SWE-fficiency open-sourced an automated coding-agent eval harness; CooperBench tests multi-agent coordination; AgentDoG traces dangerous actions to root causes.
• LM Studio 0.4: headless mode, stateful REST API, parallel requests, built-in MCP support; a hidden ROCm option suggests AMD GPU acceleration.
• Cursor pricing change: Auto mode is no longer unlimited; usage counts against the $20/month quota with token-based overage. Heavy users are shifting to Cursor CLI + terminal workflows.
• Clawdbot / Moltbot warnings: multiple communities flag the agent's centralized hosting of OpenAI/Google/Anthropic API keys as a large attack surface if compromised or prompt-injected.
• LeetCode MCP server: lets Claude log into LeetCode, fetch problems, request hints, and submit answers from the terminal; Cursor/JetBrains integrations planned.
• Prompt engineering trends: urgency phrasing ("you have 30 seconds, name one thing I missed") and micro-prompting (short imperative commands like audit/clarify/simplify) are popular — though they may trade reasoning depth for speed on deep-thinking models.

Infrastructure & Hardware

• Kimi K2.5 on dual M3 Ultra: two 512GB M3 Ultra Mac Studios over Thunderbolt 5 RDMA reach ~24 tokens/s on a quantized Kimi K2.5; Unsloth 1.8-bit dynamic quantization cuts disk usage from 600GB to 240GB.
• Decart / GPU MODE: Decart is hiring kernel engineers for Trainium 3 and the real-time video model Lucy 2; a tutorial demonstrates rolling out 1TB-class models on a single H200 via INT4 QAT + RL. Community also compares DGX (1.8TB/s bandwidth) vs RTX 5090 (300GB/s).
• tinygrad / tooling: tinygrad added an AMD simulator with DEBUG=3/6 compile/runtime instruction printing; GPU MODE members lament the lack of good interactive numeric visualization tools for quantization work.

Research & Methods

• Goodfire: mechanistic interpretability on biomedical foundation models surfaced a new class of experimentally validated Alzheimer's-related biomarkers.
• LingBot-VLA: scaling real robot manipulation data from 3K to 20K hours keeps improving VLA success rates; Qwen2.5-VL backbone plus a shared-attention action expert beats π0.5 on the GM-100 benchmark.
• MergeMix: learnable model merging mid-training can guide optimal data-mixing ratios — useful for budget-constrained teams. Paper
• Flow matching vs. diffusion vs. autoregression: community debate concludes Transformers can parameterize flow-matching vector fields, diffusion and flow matching are mathematically close, and "diffusion beats autoregression" is not settled. Flow matching | AR proposal

Products & Applications

• Gemini 3 in Chrome and Search: Gemini 3 now powers AI Overviews globally; Chrome gains sidebar integration, Nano Banana image editing, and "Auto Browse" multi-step web tasks (US Pro/Ultra preview).
• Gemini 3 Flash Agentic Vision: the model can crop, zoom, and magnify images during reasoning combined with code execution; early users report clear gains on UI/visual analysis.
• OpenAI Prism: a GPT-5.2 research workbench integrating paper reading, chart understanding, and code experiments; reception is polarized, with some criticizing it as promoting "black-box science".
• Cline 3.55.0: adds Arcee Trinity Large (400B MoE, 128K context) and Kimi K2.5 (1T MoE, 256K context, SWE-bench 76.8%), screenshot-to-UI-code with self-repair, and GPT-5 access for ChatGPT Plus/Pro subscribers without API keys.
• DeepSeek-OCR pain points: users note V1 misses headers/footers and inverted text; demand for a V2 cloud API for pipeline integration is growing.

Industry & Company News

• Kimi K2.5 business friction: ~$19/month subscription seen as pricey locally; the license requires prominent "Kimi K2.5" branding for high-DAU/high-revenue products, criticized as an adoption barrier; strong demand for availability on Perplexity and OpenRouter.
• LMArena renamed to Arena: redesigned UI resembles Claude's web app; Google login and stricter captchas frustrate users, but new sub-leaderboards (Code Arena) were added and Kimi K2.5 Thinking took open-weight #1.
• Trinity Large training cost ~$350K: per the Unsloth community, showing MoE can compress frontier-class costs, though large models still cost real money.
• Perplexity billing complaints: users report auto-renewal charges, throttled Pro query limits, and inconsistent behavior; some are escalating to banks/regulators.
• OpenRouter: launched Trinity-Large-Preview but faces refund delays of weeks to a month and unresponsive support tickets.
• Flapping Airplanes raises $180M: backers include GV, Sequoia, and Index; Karpathy commented that research startups can still outperform giants by an order of magnitude in niches.

Policy, Governance & Safety

• "Magic string" forces Claude refusals: a jailbreaking community found a token sequence that near-100% triggers refusal mode; red teamers propose standardizing such strings as emergency "circuit breakers" for high-risk trajectories.
• Safety-policy side effects: Gemini filters can be bypassed via system prompts; DeepSeek falls into repetitive refusal loops after failed jailbreaks; Chinese models show explicit filter statements in their chain of thought — evidence that most current safety layers are bolt-on rules that are both transparent and usability-damaging.
• Clawdbot/Moltbot as concentrated-key risk: not malware per se, but centralized key custody plus a exposed UI panel or prompt injection could enable bulk abuse of all stored provider keys — renewing debate over whether agents should manage keys at all versus defaulting to local, least-privilege setups.

---

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