Easy AI Daily | January 28, 2026: Kimi K2.5, Trinity Large, Transformers v5 and More

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for January 28, 2026 rounds up the day's major AI news. Moonshot released Kimi K2.5, a 1T-parameter MoE (32B active) multimodal open-weight model topping open-model benchmarks in agentic, vision, and coding tasks, with 128K→256K context and Agent Swarm (up to 100 parallel sub-agents). Arcee, Prime Intellect, and Datology launched Trinity Large, a 400B MoE (13B active) Western open-source model trained on 17T tokens. DeepSeek open-sourced DeepSeek-OCR 2; OpenAI launched Prism, a free GPT-5.2-powered research workspace; Hugging Face shipped Transformers v5 with 6–11x MoE speedups; Unsloth reported 14x MoE training acceleration. Research highlights include a proof that LLM hallucinations are inevitable, Anthropic's finding that light fine-tuning can unlock suppressed bio-risk capabilities, and Google's ATLAS multilingual scaling laws. Industry news covers Gemini Pro's real-world context shrinkage controversy, Clawdbot's rename to Moltbot over trademark and security issues, and a16z's claim that 80% of startups use Chinese open-source models.

Easy AI Daily — January 28, 2026

A structured digest of the day's AI industry news, curated by the Easy AI Daily team.

Key points

Moonshot released Kimi K2.5: an open-weight 1T-parameter MoE (32B active) multimodal model achieving open-model SOTA on HLE, BrowseComp, MMMU Pro, VideoMMMU, and SWE-bench; native image/video understanding, 128K→256K context, INT4 quantization, runs locally on multi-GPU Macs. Available on HuggingFace, Ollama, Together, Fireworks. Tech blog
Trinity Large (Arcee × Prime Intellect × Datology): 400B MoE with 13B active parameters, trained on 17T tokens (~2000 B300s for a month), featuring 3:1 local/global gated attention, SWA, NoPE+RoPE, and the Muon optimizer. Day-one vLLM support; free via OpenRouter for now. Announcement
DeepSeek-OCR 2 open-sourced: introduces Visual Causal Flow and DeepEncoder V2, compressing images to ~256–1120 vision tokens; scores 91.09% on OmniDocBench v1.5 (+3.73). Model
OpenAI Prism: a free GPT-5.2-powered research workspace combining LaTeX writing, collaboration, and literature management for all personal ChatGPT accounts; no automatic IP claims. Launch
Transformers v5 released: 6–11x MoE speedups, ~50% faster single-request inference, doubled concurrent throughput. Repo
Agent ecosystem momentum: Kimi Agent Swarm (up to 100 sub-agents, 1500 tool calls, PARL-trained scheduling), open-source Kimi Code (Apache-2.0) and Agent SDK, Google Jules' Planning Critic (-9.5% task failure), and Karpathy betting on agent-first programming.

Research highlights

A paper argues LLM hallucinations are mathematically inevitable in the current paradigm, and jailbreaks amplify them: arXiv:2409.05746
Anthropic shows minimal fine-tuning on frontier-model outputs can restore suppressed dangerous (bio-risk) capabilities: PDF
DeepPlanning benchmarks long-horizon planning; PrefixRL ~2x faster RL convergence; Google ATLAS multilingual scaling laws; Epoch's FrontierMath remains unsolved by any AI.
MergeMix uses learnable model merging to optimize data mixtures mid-training: arXiv PDF

Infrastructure & products

Unsloth: MoE training now ~14x faster than v4, targeting 30x. Post
FlagOS, tinygrad Megakernels, FlashInfer-Bench datasets released for MLSys 2026 contest: trace data
Gemini controversy: users report Pro/Ultra actual "hot memory" of only 32k–128k tokens vs. advertised millions, plus quota cuts and billing bugs (one user reportedly overcharged $70k+), pushing users to Grok 4.1 and Claude Sonnet 4.5.
Clawdbot renamed Moltbot over "Claude" trademark conflict, with community-exposed security issues around unauthorized environment-variable access.
MiniMax teases M2.2; Qwen hints at new vision models; a16z reports 80% of startups use Chinese open-source models.
AI text detectors keep mislabeling pre-GPT human papers as AI-generated, causing academic harm.

---

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