Easy AI Daily News Digest | February 12, 2026

A daily roundup of AI industry news compiled from community sources (Reddit, Twitter/X, Discord, Hugging Face).

Models & Capabilities

• Z.ai releases GLM-5 (open weights): Upgraded from 355B MoE/32B active to 744B/40B active, pretrained on 28.5T tokens, introducing DeepSeek Sparse Attention with ~200K context and long output. MIT-licensed (BF16 weights ~1.5TB). Tops the Text Arena open-model leaderboard, scores 50 on the Artificial Analysis Intelligence Index, GDPVal-AA ELO 1412, with the lowest hallucination rate in its evaluations. Links: official blog, Hugging Face, ModelScope, OpenRouter.
• GPU shortage at Z.ai: Traffic surged 10x; Z.ai prioritized Coding Plan Pro and adjusted subscription/API pricing. Reddit discussion notes even Z.ai, OpenAI, and Google face GPU supply constraints.
• GLM-5 pricing: ~$0.80/M input tokens (~3x DeepSeek V3.2 Speciale, ~1.8x Kimi K2.5); ~$2.56/M output (~6x DeepSeek). Positioned as premium quality rather than best value.
• GLM-OCR: Open-source OCR model that in some community tests beats Gemini 3 Flash; self-hostable via Hugging Face.
• DeepSeek "V4-lite" update: Chat-side support for 1M-token context, knowledge cutoff moved to May 2025, text-only. Users report strong long-document performance, but some query latency grew from 30s to 160s.
• DeepSeek's influence: Developers note nearly all frontier open models now reuse DeepSeek components—fine-grained sparse MoE, MLA, practical sparse attention, DeepEP, R1/GRPO recipes.
• MiniMax M2.5 & StepFun Step-Flash-3.5: M2.5 targets task decomposition and long execution; Step-Flash-3.5 claims #1 on MathArena. Both seen as fast, cost-effective reasoning models.
• Qwen-Image 2.0: 7B unified generation+editing model with native 2K resolution; renders complex English/Chinese typography and maintains character consistency in comics. Shrunk from 20B to 7B for local feasibility.
• Qwen3-Coder-Next on NAS iGPU: An 80B MoE run on TrueNAS SCALE + Ryzen AI iGPU via llama.cpp (Q4_K_M, Vulkan offload, flash attention) reached ~18 tok/s while running NAS and Jellyfin. Key trick: removing --no-mmap so the model fully maps into shared memory.
• Unsloth Triton MoE kernels: Claims up to 12x faster MoE training with ~35% less memory, no precision loss; works from consumer (3090) to datacenter GPUs, using the new torch._grouped_mm interface.

Agents & Tooling

• Karpathy's new paradigm: Using DeepWiki MCP + GitHub CLI, an agent read a repo, extracted needed torchao FP8 implementations into a self-contained file with tests, dropped heavy dependencies, and gained a small speedup.
• llama.cpp gains MCP support: WebUI now supports MCP tool calling over HTTP/WebSocket (GitHub, HF, Exa Search), with configurable agent loop rounds; llama-server integration planned.
• noapi-google-search-mcp: Headless Chromium wraps Google Search (web, images, reverse image search, local OCR, Lens, flights, stocks, weather, news) as an MCP tool—no official API key needed. Repo.
• mini-SWE-agent 2.0: Minimal ~100-line coding agent harness for benchmarks and RL training experiments—a shift toward auditable small harnesses over huge frameworks.
• OpenClaw: Development environment in Discord—tmux, git worktree, and Claude Code driven via channels, with context saved to markdown.
• OpenAI Codex at scale: DevRel reports ~1,500 merged PRs with no hand-written code, plus best practices for multi-hour workflows.

Infrastructure & Hardware

• $5,000 local LLM builds: Options discussed include dual 128GB Ryzen AI Max+ machines, 4×3090, 7×AMD V620, and Strix Halo; running 100B-class models with long context typically needs 40–48GB+ VRAM/unified memory.
• Multi-GPU rigs: 6×3090 (144GB VRAM) for 10B-class diffusion training; 8×3090 via x16 bifurcation; 4×4090 on X670 at reduced bandwidth (2.5GT/s x8/x4).
• CuTeDSL rising, Triton struggling on Blackwell: GPU MODE Kernelbot data shows CUDA/CuTeDSL leading submissions; Triton criticized for poor MXFP8/NVFP4 support and limited compiler control.
• torchao v0.16.0: Adds MXFP8 MoE training components with Expert Parallel support; cleanup of legacy options and progress toward ABI stability. Release notes.
• FlashInfer kernel competition: Accepts fully agent-generated and agent-assisted CUDA/CuTeDSL kernels; debate over reward hacking and whether techniques like Programmatic Dependent Launch count as cheating.

Research & Methods

• $3M open evaluation fund: Snorkel, HF, Together, Prime Intellect, Factory, Harbor, and PyTorch fund public benchmarks to close the gap between closed internal evals and public leaderboards.
• Opus 4.6 computer-use safety questioned: RedTeamCUA authors measured injection success of 83% on Opus 4.5 and ~50% on 4.6 in real Web+OS settings—far above Anthropic's system-card figures (10%/<1%)—arguing low official ASR may reflect capability limits, not safety.
• LLM "personality fingerprints": Probing 6 open 7B–9B models across 7 behavioral axes showed stable patterns (test-retest ICC 0.91–0.99) and "dead zones" RLHF may have flattened.
• Coherence over compliance: A new paper argues overtraining obedience suppresses models' latent ethical judgment, potentially making them less safe on complex dilemmas. PDF.
• Models describing inner states: Llama 3.1 / Qwen 2.5-32B invent terms like "loop" and "mirror" during self-reflection, correlating with real activation dynamics (autocorrelation r≈0.44, spectral power r≈0.62). DOI.
• Attention as optimal transport: A new paper reformulates scaled dot-product attention as a one-sided optimal transport problem. arXiv.
• DSPy community: RLM modules for memory systems, MiPROv2 for fastest-code-generation prompts, and DPO-style layering of human review signals—treating prompt engineering as a learnable optimization problem.

Products & Applications

• Seedance 2.0: Impressive text-to-video (~$0.72 per 15s by one estimate), but its photo-to-voice cloning feature was suspended by ByteDance over deepfake risks.
• SeeDance vs Veo: In a tic-tac-toe video consistency test, SeeDance sustained coherent moves ~5 steps vs Veo's 1–2.
• PixVerse R1: Real-time 720p interactive worlds—shifting from clip generation to interactive scene generation.
• AuditAI: Agentic RAG (LangGraph, Corrective RAG, strict-evidence policy) for automated NIST CSF 2.0 compliance audits with page-level citations. Code.
• Voyager VS Code extension: Converts technical papers into runnable Jupyter notebooks using Copilot.
• Control-Terminal: Open-source tool to remotely control local Claude/Codex CLI agents from a phone via Cloudflare Tunnel.

Industry & Business

• China's "agent war week": GLM-5, MiniMax M2.5, and Qwen-Image 2.0 all launched before Spring Festival, with DeepSeek V4 and Qwen 3.5 reportedly coming.
• xAI turbulence: Two co-founders, including Jimmy Ba, departed within 48 hours.
• OpenAI ads backlash: Researcher Zoë Hitzig resigned via a New York Times op-ed over plans to test ads in ChatGPT, citing privacy and behavioral targeting concerns.
• Nebius acquires Tavily: Agentic search capability folded into Nebius's AI cloud platform.
• Stripe machine payments: New billing support for AI agents as paying entities.
• Cloudflare passes $2B annual revenue: Shares rose 15%+ after hours, driven partly by AI traffic.
• DeepMind's Aletheia: Math research agent scores 91.9% on IMO-ProofBench Advanced, beating Gemini Deep Think with less compute.
• Discord ID verification backlash: Developer communities (Unsloth, Cursor, Nous) resist mandatory ID checks; some consider migrating to Matrix.

Policy, Governance & Safety

• Seedance 2.0 voice cloning halted over identity-fraud and deepfake concerns (dual-branch diffusion architecture).
• US declines to back the 2026 International AI Safety Report, drawing criticism amid a string of safety-related departures and warnings.
• Jailbreak escalation: Working jailbreaks circulate for GPT-5.2 and Gemini 3 Fast; Opus 4.6 shown easier to abuse in Google Antigravity environments. Community consensus favors hard ACL/allowlist controls over prompt-based defenses.
• Parapet multi-turn attack detection: 90.8% recall, 1.20% FPR on WildJailbreak/WildChat without LLM classifiers; fully open source. Paper, repo.

📌 Source: Easy AI Daily