Easy AI Daily | 2026-01-29

Models & Capabilities

Kimi K2.5: New Top Open Model

Moonshot's Kimi K2.5 ranked #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 at roughly 10-20% of the price. It can run locally (1T MoE; ~240GB GGUF after Unsloth quantization), though it tends to be verbose, token-heavy, and still hallucinates.

• Text Arena leaderboard
• Kimi official announcement
• Kimi-K2.5 GGUF (Unsloth)
• Local run tutorial (Reddit)

Kimi K2.5 System Prompt & Toolchain Leaked

The full system prompt and tool configuration (~5K tokens) of Kimi K2.5 were published, including a memory CRUD protocol, tool schemas, context injection, and safety rules — a valuable reference for open-source projects replicating assistant personas, memory mechanisms, and long-conversation stability.
• Leaked repo
• Kimi share link
• 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, positioned as frontier-level capability with manageable deployment cost; already available via Cline and OpenRouter.
• Arcee tech report
• OpenRouter Trinity-Large-Preview

DeepSeek-OCR 2 Released

DeepSeek's DeepSeek-OCR 2 introduces DeepEncoder V2 with a visual causal flow that first understands the page holistically before deciding reading order, significantly reducing edit distance. Community hopes V2 fixes V1's missed headers/footers and light text; users are seeking an online API for integration.
• Announcement (Reddit)
• Online demo

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

Alibaba Tongyi released Z-Image Base and Turbo. Base uses more steps for refined quality; Turbo generates in ~7 seconds with a more photorealistic look. Community tests show it runs on 12GB VRAM; the Stable Diffusion community calls it "something like a next-gen SDXL."
• Z-Image model page
• Base vs Turbo discussion

DeepMind Open-Sources AlphaGenome

Google DeepMind released AlphaGenome weights for predicting the molecular impact of genetic variants, after 1M+ daily API calls and 3,000+ users; weights are on Hugging Face.
• Announcement

NVIDIA NVFP4 Nemotron 3 Nano

NVIDIA shipped an NVFP4-precision Nemotron 3 Nano claiming 4x BF16 throughput on Blackwell B200 with 99.4% accuracy retained via quantization-aware distillation (QAD). vLLM already supports the format.
• NVIDIA AI Dev
• vLLM NVFP4 support

MongoDB LEAF: Embeddings Distilled for CPU

MongoDB Research's LEAF computes embeddings offline with a large model and serves queries with a small aligned model: ~96% quality retained, 5-15x smaller models, up to 24x query throughput, runnable on CPU/edge devices.
• LEAF overview

Frontier Model "Personalities": GPT vs Claude

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

Agents & Tooling

• Agent Skills emerging: DeepLearning.AI + Anthropic launched an Agent Skills course (extracting workflow logic into reusable skill files); LangChain is pushing Skills; Hugging Face's upskill shows strong-model trajectories boosting weak models on CUDA kernel tasks by up to 45% — results vary by model. (Andrew Ng, LangChain, HF upskill)
• Trajectory-based agent evals: SWE-fficiency released an automated coding-agent eval harness; CooperBench targets multi-agent collaboration; AgentDoG traces dangerous actions to root causes across trajectories. (SWE-fficiency, CooperBench)
• LM Studio 0.4: headless mode, stateful REST API, parallel requests, built-in MCP support; hidden ROCm options suggest AMD GPU acceleration. (Release blog)
• Cursor pricing changes: Auto mode no longer unlimited; usage counts against the $20/month quota with token billing beyond, frustrating heavy users; some shift to Cursor CLI for large projects.
• Clawdbot / Moltbot security warning: communities warn that centralizing OpenAI/Google/Anthropic API keys under one agent creates a pool vulnerable to compromise or prompt injection.
• LeetCode MCP server: lets Claude log into LeetCode, fetch problems, get hints, and submit solutions from the terminal. (GitHub)
• Prompt engineering tricks: time-pressure phrasing ("you have 30 seconds...") and micro-prompting (short command+verb calls) are popular; caveat: 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 run quantized Kimi K2.5 at ~24 tok/s; Unsloth 1.8-bit dynamic quantization cuts disk usage from 600GB to 240GB. (Test)
• GPU MODE / Decart: Decart is hiring kernel engineers for Trainium 3 and real-time video model Lucy 2; a tutorial shows deploying 1TB-class models on a single H200 via INT4 QAT+RL; DGX vs RTX 5090 discussions highlight memory bandwidth (1.8TB/s vs ~300GB/s) as the real bottleneck.
• tinygrad / tooling: tinygrad added an AMD simulator (DEBUG=3/6) for kernel debugging without hardware; GPU MODE members note the lack of good interactive numerical-visualization tools for quantization work. (tinygrad PR #14387)

Research & Methods

• Goodfire: mechanical interpretability on biomedical foundation models surfaced a new class of experimentally validated Alzheimer's-related biomarkers — a reusable "superhuman model + interpretability" pipeline for scientific discovery. (Thread)
• LingBot-VLA: scaling real robot manipulation data from 3k to 20k hours keeps improving VLA success rates; the architecture uses Qwen2.5-VL plus an attention-sharing action expert, beating π0.5 on GM-100. (Analysis)
• MergeMix: learnable mid-training model merges guide optimal data-mix ratios — useful for budget-constrained open-source teams. (Paper)
• Flow matching vs diffusion vs autoregression: community consensus — Transformers can parameterize flow matching vector fields; diffusion and flow matching are mathematically close; "diffusion beats autoregression" isn't established, many gaps are architectural compromises. (Flow matching, AR proposal)

Products & Applications

• Gemini 3 in Chrome and Search: now powers AI Overviews globally; Chrome gains sidebar integration, Nano Banana image editing, and "Auto Browse" multi-step web tasks (US Pro/Ultra preview). Developers call it the most agent-like browser integration yet. (Chrome thread)
• Gemini 3 Flash Agentic Vision: the model can actively crop, zoom, and enlarge image regions during reasoning, combined with code execution; front-end engineers report clear gains in UI/visual analysis. (Blog)
• OpenAI Prism: a GPT-5.2 research workspace integrating paper reading, chart understanding, and code experiments; reception is polarized, with some criticizing "black-box science." (Bubeck response)
• Cline 3.55.0: adds Arcee Trinity Large (128K context) and Kimi K2.5 (256K context, SWE-bench 76.8%), with screenshot-to-UI-code and self-repair; GPT-5 series available to ChatGPT Plus/Pro users without an API key. (Changelog)

Industry & Business

• Kimi K2.5 licensing/pricing: ~$19/month subscription seen as pricey locally; license requires prominent "Kimi K2.5" branding and extra limits for high-DAU/high-revenue products, criticized as a barrier to enterprise adoption; strong community demand for Perplexity/OpenRouter listings.
• LMArena renamed to Arena: redesigned UI resembling Claude's web app, Google login, stricter CAPTCHAs — controversial; adds Code Arena and other sub-leaderboards where Kimi K2.5 Thinking ranks #1 among open models. (Blog)
• Trinity Large training cost: reportedly ~$350K — evidence MoE can cut frontier-adjacent training costs, but big models still cost real money.
• Perplexity billing complaints: Discord users report auto-renewals with degraded service, fluctuating query limits, and plans to escalate to banks/regulators.
• OpenRouter: added Trinity/Kimi models but faces refund delays of weeks and slow support responses.
• Flapping Airplanes raises $180M: backed by GV, Sequoia, Index; Karpathy notes new research startups can still beat incumbents by an order of magnitude in niches.

Policy, Governance & Safety

• Claude "magic string" refusal: a special token sequence forces near-100% refusal mode; red-teamers propose standardizing such strings as "circuit breakers" for high-risk trajectories.
• Safety policy side effects: Gemini filters bypassable via system prompts; DeepSeek can loop endlessly in refusals after failed jailbreaks; visible filtering in some Chinese models' chains of thought — suggesting current safety is largely a bolt-on layer that's both transparent and usability-damaging.
• Clawdbot/Moltbot as centralized-key risk: not malware per se, but an exposed panel or prompt injection could batch-invoke all stored cloud API keys — sparking debate over whether agents should manage keys at all vs. local-first, least-privilege defaults.

---

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