Easy AI Daily News | January 28, 2026: Kimi K2.5, Trinity Large, DeepSeek-OCR 2, and More

📅 AI Industry News — January 28, 2026

Models & Capabilities

#### Moonshot Releases Kimi K2.5: Open-Source Multimodal MoE SOTA Kimi K2.5 uses a 32B-activated / 1T-parameter MoE architecture, topping open-source leaderboards on agentic, vision, and coding benchmarks including HLE, BrowseComp, MMMU Pro, VideoMMMU, and SWE-bench. It natively supports image + video understanding, can reconstruct web pages from screen recordings, supports 128K→256K context, and INT4 partial quantization. Available on HuggingFace, Ollama, Together, and Fireworks — some setups run locally on multi-GPU Macs.

• Official blog | Moonshot release thread | Zach Mueller technical breakdown | LMArena leaderboard | Local run on M3 Ultra

#### Arcee / Prime Intellect Release Trinity Large: A 400B MoE Western Open-Source Counterattack Trinity Large (preview): 400B MoE, 13B activated, trained on 17T tokens. Features 3:1 local/global gated attention, SWA, NoPE+RoPE, depth-scaled normalization, and the Muon optimizer. Trained for ~1 month on ~2000 B300 GPUs with Arcee, Prime Intellect, and Datology. vLLM offered day-one support; OpenRouter is temporarily free.

• Arcee announcement | Prime Intellect intro | Technical highlights | vLLM support

#### DeepSeek-OCR 2: Document OCR That Learns "Reading Order" Open-sourced on HuggingFace with Visual Causal Flow and DeepEncoder V2. Compresses images into ~256–1120 vision tokens; scores 91.09% (+3.73) on OmniDocBench v1.5. Community reviewers note it's below real SOTA but the approach is worth watching. vLLM inference supported.

• Model page | vLLM support | Jerry Liu's take

#### OpenAI Prism: A GPT-5.2-Powered "Research Overleaf" OpenAI launched Prism, a free research workspace integrating LaTeX writing, collaboration, and reference management, powered by GPT-5.2 for rewriting, citations, and search. Available to all personal ChatGPT accounts; data use follows existing ChatGPT policies with no automatic claim on research IP.

• OpenAI launch | Data/IP clarification

#### Qwen Updates: Thinking Models & Quantization Qwen3-Max-Thinking claims reasoning close to commercial closed models, though community reports code-agent compile issues. Qwen3-32B INT4 quantization on H100: only ~1.9% MMLU-Pro accuracy loss, while concurrent users jump from 4 to 47 (4k context).

• Qwen3-Max-Thinking blog | Quantization evaluation

#### Transformers v5 Stable: MoE Speedups at Scale Transformers v5 delivers 6–11x MoE prompt performance gains, simplified tokenizer/backend config, faster dynamic weight loading, and supports quantization + MoE + parallelism + PEFT. Community tests: ~50% faster single-request inference, 2x concurrent throughput.

• Repo & migration guide

---

Agents & Tooling

#### Kimi Agent Swarm: Up to 100 Parallel Sub-Agents Built into K2.5 (paid beta), Agent Swarm dynamically spawns up to 100 sub-agents with up to 1500 tool calls, cutting end-to-end task time to ~1/4–1/3 of a single agent. Trained via Parallel Agent RL (PARL).

• Agent Swarm explained | Technical report summary

#### Kimi Code & Agent SDK Moonshot released Kimi Code, an open-source (Apache-2.0) coding agent with IDE integration, plus an Agent SDK for custom workflows.

• Kimi Code | Agent SDK

#### Multi-Sub-Agent + Planning Critic Becomes a Consensus Pattern Kimi Swarm, LangChain sub-agents, and Google Jules' Planning Critic share an architecture: an orchestrator splits tasks across parallel sub-agents, with a "plan reviewer" critiquing before execution. Jules claims ~9.5% lower task failure rate.

• LangChain sub-agents | Jules Planning Critic

#### IDE Agents: Cursor, VS Code MCP Cursor reports significantly better agent quality after semantic indexing of large codebases. VS Code is improving "command execution explanation" UX, and MCP Apps lets tool servers return UI components directly.

• Cursor update | MCP Apps UI example

#### Claude "Hive" Multi-Agent Coding Experiment A community project built 7 specialized Claude Code agents (coding, testing, review) sharing SQLite+FTS5 memory via MCP, compared against Microsoft Autogen and BMAD. CheshireCat Enterprise focuses on multi-tenant agent workflows.

• Claude multi-agent project | Microsoft Autogen | CheshireCat core

#### Karpathy Bets on Agent-First Programming Andrej Karpathy says he's migrating his workflow to LLM agents (Claude et al.) that continuously code and run experiments while humans supervise — "models never get tired."

• Karpathy post

---

Infrastructure & Hardware

#### Unsloth: 14x MoE Training Speedup, Targeting 30x With new kernels and Transformers v5, Unsloth reports ~14x faster MoE training vs v4, aiming for 30x, targeting low-cost training of large MoE models on consumer/cloud GPUs.

• Announcement

#### FlagOS: Unifying Model–System–Chip GPU MODE community proposes FlagOS, an open-source system stack spanning models, systems, and chips for portable, co-designed AI workloads.

• Discussion thread

#### Tinygrad & Megakernels: From FlashAttention to a "GPU OS" tinygrad can now auto-rewrite naive attention frontends into FlashAttention kernels and is exploring compiling models into monolithic Megakernels. George Hotz cites Luminal: future GPUs resemble an OS running one big kernel.

• Megakernel post

#### Multi-GPU / Used-GPU Local LLM: Bandwidth & Cooling Are the Real Bottlenecks Reddit/Eleuther discussions on stacking used Teslas to 200GB+ VRAM: bottlenecks are PCIe bandwidth and prefill speed rather than token/s, plus cooling and power costs.

• Benchmarks & discussion

#### FlashInfer-Bench / MLSYS26 Datasets Released FlashInfer published inference traces and MLSys 2026 contest workloads on HuggingFace, with biweekly leaderboards planned for optimizing decoding and MoE kernels.

• Trace dataset | Contest dataset

---

Research & Methods

#### Math Proof: LLMs Will Always Hallucinate — Jailbreaking Makes It Worse A paper proves hallucination is unavoidable in current paradigms; jailbreaks exploit this by shifting context so models no longer flag malicious content, amplifying hallucination and safety risks.

• On the Inevitable Hallucinations of LLMs

#### Anthropic: Minimal Fine-Tuning "Unlocks" Suppressed Capabilities Anthropic shows light fine-tuning of open models on closed-model outputs can restore or enhance suppressed dangerous capabilities (e.g., bio risk) with modest compute — evidence that refusal-only safety is fragile and open-model dual-use risk is underestimated.

• Paper PDF | Official tweet

#### DeepPlanning & PrefixRL DeepPlanning benchmarks long-horizon verifiable tasks (multi-day travel, shopping) — current agents still struggle. PrefixRL reuses old trajectories via prefix conditioning, ~2x faster RL convergence for complex reasoning.

• DeepPlanning | PrefixRL

#### Multilingual Scaling Laws & FrontierMath Google Research's ATLAS provides multilingual scaling-law guidance; Epoch's FrontierMath: Open Problems shows no AI has solved any included genuine research math problems yet.

• ATLAS | FrontierMath

#### MergeMix: Learnable Model Merging for Data Mixing MergeMix searches optimal data mixtures via learnable model merging mid-training — practical for compute-constrained open-source teams.

• Paper

---

Products & Applications

• Kimi K2.5 Office Agent: widely used in China for report writing and multi-document synthesis; some companies replaced weekly-report/analysis workflows with it. Blog
• Gemini AI Studio downgraded: free quotas cut; Pro/Ultra "hot memory" reportedly only 32k–128k tokens despite million-token marketing, pushing users toward Grok 4.1 (2M) and Claude Sonnet 4.5 (1M) or vector retrieval. Discussion
• Perplexity Pro friction: rate-limited search/image generation, opaque billing, and payment failures (notably India) drive some users to privacy-focused Kagi.
• Local coding assistants: on 8GB VRAM + 32GB RAM, Qwen2.5/3 Coder 7B/30B is the practical ceiling; Cline-style multi-agent plugins often hit CUDA OOM without careful tuning.

---

Industry & Company News

• Chinese open-source momentum: Kimi K2.5 rivals Claude Opus 4.5 and Gemini 3 Pro on several benchmarks with open weights; a16z claims 80% of startups use Chinese open-source models. Artificial Analysis
• Pre-holiday releases: MiniMax teases M2.2; Qwen hints at new vision/multimodal models (Z-Image). MiniMax M2.2
• Trinity Large as an "American return" to from-scratch pretraining rather than post-training only. Discussion
• Clawdbot renamed Moltbot under Anthropic trademark pressure; worse, community reports it reads environment variables/secrets without authorization — multiple communities advising against use. Rename | Security video
• Decart releases Lucy 2, an autoregressive video editing model, and is hiring kernel engineers for real-time video/world models on Trainium 3. Report

---

Policy, Governance & Safety

• AI detectors flagging pre-GPT papers: OpenAI Discord reports human-written pre-LLM papers judged "AI-generated"; accuracy is poor yet institutions keep using them.
• Gemini Pro context shortfall: users measured ~32k hot context vs advertised 1–2M, plus billing bugs (one erroneous charge reportedly over $70k) damaging developer trust. Thread
• GPT-5 "control shell" leak: a file named GPT5_Hotfix.md, claimed to be a pre-generation control shell with grammar constraints, intent locking, and drift protection — if genuine, frontier labs increasingly rely on outer rule layers. Screenshot

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