Easy AI Daily | January 28, 2026

A curated digest of AI industry news from the Easy AI Daily forum post on zhichai.net, covering models, agents, infrastructure, research, products, and policy.

Models & Capabilities

Moonshot Releases Kimi K2.5: Open-Source Multimodal MoE SOTA

• Kimi K2.5 uses a 32B-active / 1T-parameter MoE architecture, ranking first among open models on agentic, vision, and coding benchmarks including HLE, BrowseComp, MMMU Pro, VideoMMMU, and SWE-bench.
• Native image + video understanding; can reconstruct webpages 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.
• Links: Official blog | Release thread | LMArena open-model rankings

Arcee / Prime Intellect Release Trinity Large

• Preview of a 400B MoE with 13B active parameters, trained on 17T tokens using ~2000 B300 GPUs for a month.
• Architecture: 3:1 local/global gated attention, SWA, NoPE+RoPE, depth-scaled normalization, Muon optimizer.
• vLLM day-one support; temporarily free via OpenRouter.
• Links: Arcee announcement | Prime Intellect intro

DeepSeek-OCR 2: OCR That Learns Reading Order

• Open-sourced on HuggingFace with Visual Causal Flow and DeepEncoder V2; compresses images to ~256–1120 vision tokens.
• OmniDocBench v1.5 score: 91.09% (+3.73). vLLM inference already supported.
• Links: Model page | vLLM support

OpenAI Prism: GPT-5.2-Powered Research Workspace

• Free workspace integrating LaTeX writing, collaboration, and reference management for all personal ChatGPT accounts.
• Data usage follows existing ChatGPT policies; no automatic IP claims on research output.
• Links: OpenAI release | Data/IP clarification

Qwen Updates

• Qwen3-Max-Thinking claims near-closed-model reasoning, though users report compile issues in code-agent mode.
• Qwen3-32B INT4 quantization on H100: only ~1.9% MMLU-Pro accuracy loss while concurrency rose from 4 to 47 users (4k context).
• Links: Qwen3-Max-Thinking blog | Quantization evaluation

Hugging Face Transformers v5

• 6–11x MoE prompt performance improvements, 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: Repo and migration guide

Agents & Tooling

Kimi Agent Swarm

• Beta for paid users: dynamically spawns up to 100 sub-agents, up to 1500 tool calls, claiming end-to-end task time reduced to ~1/4–1/3 of a single agent. Trained via Parallel Agent Reinforcement Learning (PARL).

Kimi Code & Agent SDK

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

Multi-Sub-Agent + Planning Critic Patterns

• LangChain sub-agents and Google Jules' Planning Critic converge on the same architecture: an orchestrator decomposes tasks for parallel sub-agents, with a planning reviewer critiquing before execution. Jules claims ~9.5% reduction in task failure rate.

IDE Agents

• Cursor reports better agent quality/speed after semantic indexing of large codebases; VS Code improves command-execution safety UX and MCP Apps returning UI components directly.

Karpathy Bets on Agent-First Programming

• Andrej Karpathy is migrating his workflow to LLM agents (e.g., Claude) that continuously code and run experiments, with humans supervising — citing "models never get tired" as the key advantage.

Infrastructure & Hardware

• Unsloth: MoE training now ~14x faster than v4 with new kernels and Transformers v5; targeting 30x. (announcement)
• FlagOS: GPU MODE community proposal for an open system stack spanning models–systems–chips for portable, tunable AI workloads.
• Tinygrad/Megakernels: auto-rewriting naive attention into FlashAttention kernels; compiling models into monolithic megakernels. See Luminal blog.
• Multi-GPU / older GPUs for local LLMs: bottlenecked by PCIe bandwidth and prompt prefill rather than token/s; cooling and power costs often not worth it. (benchmark)
• FlashInfer-Bench / MLSYS26: inference trace and contest datasets released on HuggingFace with a biweekly leaderboard. (trace dataset | contest dataset)

Research & Methods

• Hallucinations are provably inevitable: A paper theoretically proves LLMs cannot fully eliminate hallucinations under the current paradigm; jailbreaks exploit this by shifting context so models no longer flag malicious/unreliable content. (arXiv:2409.05746)
• Anthropic bio-risk paper: light fine-tuning of open models on frontier-model outputs can restore or amplify suppressed dangerous capabilities at low compute cost — suggesting refusal-only safety is fragile. (PDF)
• DeepPlanning benchmarks long-horizon verifiable tasks (multi-day trips, shopping), showing agents still struggle; PrefixRL reuses old trajectories via prefix conditioning for ~2x faster RL convergence.
• ATLAS (Google Research): multilingual scaling-law guidance on data mixtures and model size; FrontierMath: Open Problems: no AI has yet solved the included genuine math research problems.
• MergeMix: learnable model merging to search data-mix ratios mid-training, useful for compute-limited open projects. (PDF)

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.
• Gemini AI Studio downgrades: free quotas cut; Pro/Ultra reportedly use only 32k–128k token "hot memory" despite million-token marketing, pushing users to Grok 4.1 (2M) or Claude Sonnet 4.5 (1M), or toward retrieval instead of long contexts. (discussion)
• Perplexity Pro vs Kagi: rate limits, opaque billing, and payment failures (notably in India) are pushing some users toward privacy-focused Kagi.
• Local coding assistants: on 8GB VRAM + 32GB RAM machines, Qwen Coder 7B/30B is the practical ceiling; Cline-style multi-agent plugins often hit CUDA memory errors requiring manual tuning.

Industry & Company News

• Chinese open models vs Western closed models: Kimi K2.5 rivals or beats Claude Opus 4.5 and Gemini 3 Pro on multiple benchmarks while being open-weight; a16z reports 80% of startups use Chinese open-source models. (Artificial Analysis)
• MiniMax M2.2 teaser and Qwen hints at new vision/multimodal models (e.g., Z-Image) — typical pre-holiday release cadence.
• Trinity Large as a Western return to from-scratch pretraining rather than fine-tuning and evaluation only.
• Clawdbot renamed Moltbot due to Claude trademark conflict; also facing security reports of unauthorized access to environment variables and secrets.
• Decart Lucy 2: autoregressive video-editing model; hiring kernel engineers for real-time video/world models on accelerators like Trainium 3. (technical report)

Policy, Governance & Safety

• AI detectors flagging pre-GPT papers as AI-generated: poor accuracy but still used by universities and recruiters.
• Gemini Pro context-shortfall accusations: measured ~32k effective hot context vs advertised 1–2M; plus billing bugs (some users wrongly charged over $70,000) damaging developer trust.
• GPT-5 "control shell" leak: a file named GPT5_Hotfix.md circulating in jailbreak communities claims to be a pre-generation control layer enforcing syntax constraints, intent locking, and drift prevention — if genuine, frontier labs increasingly rely on external rule layers.

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