Easy AI Daily | January 28, 2026

A digest of AI industry news covering models, agents, infrastructure, research, products, and policy.

Models and Capabilities

Moonshot releases Kimi K2.5: open-source multimodal MoE SOTA

• Kimi K2.5 uses a 32B-active / 1T-parameter MoE architecture and takes first place among open models on agent, vision, and coding benchmarks including HLE, BrowseComp, MMMU Pro, VideoMMMU, and SWE-bench.
• Native image + video understanding; can reconstruct web pages from screen recordings; supports 128K→256K context and INT4 partial quantization.
• Available on HuggingFace, Ollama, Together, Fireworks; can run locally on multi-GPU Macs in some configurations.
• Links: Official blog | Announcement | Technical breakdown | LMArena leaderboard | Local run on M3 Ultra

Arcee / Prime Intellect release Trinity Large: a Western open-source counterattack

• Trinity Large preview: 400B MoE with 13B active parameters, trained on 17T tokens. Features 3:1 local/global gated attention, SWA, NoPE+RoPE, depth-scaled normalization, Muon optimizer. Trained on ~2000 B300 GPUs for about a month.
• Day-one vLLM inference support; temporarily free via OpenRouter.
• Links: Arcee announcement | Prime Intellect | Technical summary | vLLM support

DeepSeek-OCR 2: document OCR that learns "reading order"

• Open-sourced on HuggingFace with Visual Causal Flow and DeepEncoder V2; compresses images to ~256–1120 vision tokens; scores 91.09% on OmniDocBench v1.5 (+3.73).
• Community assessment: below true SOTA but a promising direction; vLLM inference supported.
• Links: Model page | vLLM | Jerry Liu analysis

OpenAI Prism: a GPT-5.2-powered "research Overleaf"

• Free research workspace combining LaTeX writing, collaboration, and literature management; open to all ChatGPT personal accounts. Data use follows existing ChatGPT policies; no automatic IP claims on research output.
• Links: Announcement | Product lead notes

Qwen: thinking models and quantization

• Qwen3-Max-Thinking claims near-frontier reasoning, though community reports compilation issues in its coding agent mode.
• Qwen3-32B INT4 quantization on H100: only ~1.9% MMLU-Pro accuracy loss while concurrent users rose from 4 to 47 (4k context).
• Links: Qwen3-Max-Thinking | Quantization evaluation

Transformers v5 stable release

• 6–11x MoE speedups, simplified tokenizer/backend config, faster dynamic weight loading, support for quantization + MoE + parallelism + PEFT. Community tests: ~50% faster single-request inference, doubled concurrent throughput.
• Link: Repository & migration guide

Agents and Tooling

Kimi Agent Swarm: up to 100 parallel sub-agents

• Beta for paid users; dynamically spawns up to 100 sub-agents with up to 1500 tool calls; claims end-to-end task time down to ~1/4–1/3 of a single agent. Trained via parallel agent RL (PARL).
• Links: Explanation | Technical report

Kimi Code and Agent SDK

• Kimi Code: open-source coding agent (Apache-2.0) with IDE integration; Agent SDK for building custom agent workflows.
• Links: Kimi Code | Agent SDK

Multi-agent + planning-critic patterns

• Kimi Swarm, LangChain sub-agents, and Google Jules' Planning Critic converge on one orchestrator splitting tasks among parallel sub-agents plus a plan reviewer. Jules claims ~9.5% lower task failure rate.
• Links: LangChain | Jules Planning Critic

IDE agents: Cursor, VS Code MCP

• Cursor: semantic search indexing of large codebases improves agent quality and latency; VS Code improving command-execution safety UX; MCP Apps let tool servers return UI components.
• Links: Cursor | MCP Apps example

Multi-agent coding setups

• A community-built "hive" of 7 specialized Claude Code agents sharing SQLite+FTS5 memory via MCP, compared against Microsoft Autogen and BMAD. CheshireCat Enterprise targets multi-tenant agent workflows.
• Links: Claude multi-agent project | Autogen

Karpathy bets on "agent-first programming"

• Andrej Karpathy says he's shifting his workflow to LLM agents (e.g., Claude) that continuously code and run experiments while humans supervise, citing "models never get tired" as the key advantage.
• Link: Post

Infrastructure and Hardware

• Unsloth: MoE training ~14x faster than v4 with new kernels + Transformers v5, targeting 30x. Announcement
• FlagOS: GPU MODE community proposal for an open stack unifying model–system–chip layers. Discussion
• Tinygrad / Megakernels: auto-rewriting attention frontends into FlashAttention kernels; compiling models into large unified megakernels; George Hotz cites Luminal's "GPU as an OS" thesis. Blog post
• Multi-GPU local LLM on used hardware: benchmarks with ~200GB+ VRAM Tesla rigs show PCIe bandwidth and prompt prefill—not token/s—as the bottleneck; cooling/power often not worthwhile. Benchmarks
• FlashInfer-Bench / MLSYS26: inference trace datasets and contest workloads released on HuggingFace, with biweekly leaderboards planned. Traces | Contest data

Research and Methods

• Hallucinations are provably inevitable: a paper discussed in BASI argues LLMs cannot fully eliminate hallucinations under the current paradigm; jailbreaks exploit this by shifting context so models no longer flag unreliable content. arXiv:2409.05746
• Anthropic bio-risk paper: minor fine-tuning of open models on frontier-model outputs can restore or amplify capabilities suppressed by safety training, at low compute cost—showing refusal-only safety is fragile. PDF
• DeepPlanning & PrefixRL: verifiable-constraint long-horizon planning benchmarks show agents still struggle with real-world plans; prefix-conditioned trajectory reuse roughly doubles RL convergence speed for complex reasoning. DeepPlanning | PrefixRL
• Multilingual scaling laws & FrontierMath: Google's ATLAS provides data-mix/size scaling guidance for multilingual LLMs; Epoch's FrontierMath: Open Problems remains unsolved by any AI, calibrating expectations of "AI researchers." ATLAS | FrontierMath
• MergeMix: learnable model merging to search for better data mixtures mid-training, aimed at compute-constrained open projects. Paper

Products and Applications

• Kimi K2.5 office agent: widely used for report writing and document synthesis; Moonshot highlights end-to-end multi-document retrieval-to-draft automation, with some companies replacing weekly-report workflows. Blog
• Gemini AI Studio downgrades: free-tier cuts and reports that Pro/Ultra "hot memory" is effectively only 32k–128k tokens despite million-token marketing; users migrating to Grok 4.1 (2M) and Claude Sonnet 4.5 (1M) or switching to RAG. Discussion
• Perplexity Pro vs. Kagi: rate limiting, opaque billing (notably failed payments in India) push some users toward Kagi's privacy-focused, Claude-integrated search.
• Local coding assistants: on 8GB VRAM + 32GB RAM machines, Qwen2.5/3 Coder 7B/30B is the practical ceiling; Cline-style multi-agent plugins often hit CUDA OOM errors requiring context-length tuning—plug-and-play "Copilot-grade" local experience remains distant.

Industry and Company News

• Chinese open-source momentum: Kimi K2.5 approaches or beats Claude Opus 4.5 and Gemini 3 Pro on several agentic/coding/multimodal benchmarks with open weights; a16z reports ~80% of startups using Chinese open-source models. Artificial Analysis
• Pre-holiday releases: MiniMax teases M2.2; Qwen hints at new vision/multimodal models (Z-Image) via ComfyUI PRs—classic "pre-Lunar New Year" release cadence. MiniMax teaser
• Trinity Large as a Western reboot: seen as the first serious from-scratch large-scale pretraining from a Western open-source group in a while, with data from Datology and compute from Prime Intellect.
• Clawdbot → Moltbot: renamed due to Anthropic trademark conflict; community also disclosed unauthorized reads of environment variables/secrets, prompting several communities to advise against use. Rename | Security video
• Decart Lucy 2: autoregressive video editing model released; hiring engineers for low-latency kernels on accelerators like Trainium 3. Report

Policy, Governance, and Safety

• AI detectors flag pre-GPT papers: OpenAI Discord and elsewhere report AI-text detectors misclassifying papers written before GPT existed; institutions and recruiters still rely on them.
• Gemini Pro context controversy: experiments suggest ~32k effective hot context (far below claimed 1–2M), plus limit cuts and billing bugs (one user allegedly overcharged by over $70,000) damage Google's developer trust. Thread
• GPT-5 "control shell" leak: a file named GPT5_Hotfix.md posted in BASI Jailbreaking claims to be GPT-5's pre-generation control layer enforcing syntax constraints, intent locking, and drift protection—suggesting frontier labs increasingly rely on external rule layers.

📌 Source: Easy AI Daily