Easy AI Daily Digest | February 12, 2026: GLM-5, Chinese Model Releases, Agent Tooling, and AI Policy News

Forum topic · 小凯 · 2026-03-27

Summary

A comprehensive AI industry digest from February 12, 2026, covering major model releases, agent tooling, infrastructure developments, research, and policy news. Highlights include Zhipu's Z.ai releasing GLM-5 as open-weight MIT-licensed model (744B MoE, 40B active parameters, ~200K context), DeepSeek's reported V4-lite update with 1 million token context, Alibaba's Qwen-Image 2.0 (7B unified generation/editing model), and MiniMax M2.5. The agent ecosystem saw Karpathy demonstrating agent-driven code extraction workflows, MCP support in llama.cpp, and OpenAI showcasing 1,500 PRs merged via Codex. Hardware discussions cover $5,000 local LLM setups and Unsloth's Triton MoE kernels claiming 12x training speedups. Industry news includes xAI co-founder departures, an OpenAI researcher resigning over ad plans, Nebius acquiring Tavily, Stripe launching payments for AI agents, and ByteDance suspending Seedance 2.0's voice cloning feature over deepfake concerns. The digest reflects an intensifying 'Chinese model release week' that Western communities describe as a systematic price-and-performance challenge to closed-lab offerings.

Key points

Models and Capabilities

Z.ai releases GLM-5: Upgraded from 355B MoE/32B active to 744B/40B active parameters, pretrained on 28.5T tokens with DeepSeek Sparse Attention, ~200K context. Open-sourced under MIT license (~1.5TB BF16 weights). Tops Text Arena among open models; Artificial Analysis intelligence index score of 50; lowest hallucination rate in its evaluations. Official blog: https://z.ai/blog/glm-5, weights on Hugging Face and ModelScope.
GLM-5 positioning and compute constraints: Z.ai revealed that 'Pony Alpha' is GLM-5, focused on agent scenarios and long tasks. Traffic surged 10x, causing GPU shortages; only Coding Plan Pro is prioritized and pricing was adjusted. Community notes even Z.ai, OpenAI, and Google are compute-constrained.
GLM-5 pricing: ~$0.80/M input tokens (about 3x DeepSeek V3.2 Speciale, 1.8x Kimi K2.5) and ~$2.56/M output tokens (about 6x DeepSeek). Positioned as premium quality rather than cost-efficiency.
GLM-OCR: Open-source small OCR model that outperforms Gemini 3 Flash on some OCR tasks; self-hostable from https://huggingface.co/zai-org/GLM-OCR.
DeepSeek 'V4-lite': Reports of a new version supporting 1M token context, knowledge cutoff updated to May 2025, text-only. Users report strong long-document performance, but latency rose from 30s to 160s on some queries.
DeepSeek's influence on open-source: Fine-grained sparse MoE, MLA, sparse attention practicalization, DeepEP, and the R1/GRPO recipe are now reused by nearly all frontier open models including GLM-5.
MiniMax M2.5 and StepFun Step-Flash-3.5: Both positioned as cost-efficient reasoning models; Step-Flash-3.5 claims #1 on MathArena.
Qwen-Image 2.0: Alibaba's 7B unified image generation/editing model with native 2K resolution, complex English/Chinese typography rendering, and multi-panel comic character consistency—downsized from 20B to 7B for local-run potential.
Qwen3-Coder-Next 80B on an NAS iGPU: A user ran the 80B MoE model on TrueNAS SCALE with a Ryzen AI iGPU at ~18 tok/s using llama.cpp, Q4_K_M quantization, Vulkan offload, and flash attention; the key fix was removing --no-mmap to map the model into shared memory.
Unsloth Triton MoE kernels: Claims up to 12x faster MoE training with ~35% less memory, no precision loss, working on consumer GPUs (including 3090) via torch._grouped_mm.

Agents and Tooling

Karpathy's new paradigm: Used DeepWiki MCP + GitHub CLI to let an agent read a repo, then extract needed torchao FP8 implementations into a self-contained file with tests, dropping heavy dependencies with a small speedup.
llama.cpp adds MCP support: WebUI now supports configurable agent loops and remote MCP servers (GitHub, HF, Exa Search) over HTTP/WebSocket; llama-server integration planned.
noapi-google-search-mcp: Headless Chromium wraps Google Search as an MCP tool without API keys—web, images, reverse image search, OCR, flights, stocks, weather, and news for local models. Repo: https://github.com/VincentKaufmann/noapi-google-search-mcp
mini-SWE-agent 2.0: A ~100-line minimal coding agent for benchmarking and RL experiments, reflecting a shift toward auditable small harnesses over huge frameworks.
OpenClaw: Chaining tmux, git worktree, and Claude Code to develop entirely from Discord with persistent markdown context.
OpenAI Codex case study: DevRel reports ~1,500 PRs merged with no handwritten code, plus best practices for multi-hour agent workflows.

Infrastructure and Hardware

$5,000 local LLM builds (2026): Community options include dual 128GB Ryzen AI Max+ machines, 4x3090 or 7x AMD V620, and Strix Halo silent builds. Consensus: 100B-class long-context inference needs 40–48GB VRAM/unified memory minimum.
Multi-GPU rigs: 6x3090 (144GB) for diffusion training, 8x3090 via x16 bifurcation, and 4x4090 on X670 boards—with bandwidth and power-stability caveats.
CuteDSL vs Triton on Blackwell: GPU MODE Kernelbot data shows CUDA and CuTeDSL lead submissions; Triton criticized for poor MXFP8/NVFP4 support on Blackwell.
torchao v0.16.0: Adds MXFP8 MoE training components, Expert Parallel support, cleanup of legacy quantization options, and progress toward ABI stability. https://github.com/pytorch/ao/releases/tag/v0.16.0
FlashInfer AI kernel competition: Categories for fully agent-generated and agent-assisted CUDA/CuTeDSL kernels; community debate on preventing reward hacking and rules on advanced techniques like Programmatic Dependent Launch.

Research and Methods

$3M open evaluation fund: Snorkel, Hugging Face, Together, Prime Intellect, Factory, Harbor, and PyTorch fund open benchmarks to close the gap between closed-lab internal scores and public leaderboards.
Opus 4.6 computer-use safety dispute: RedTeamCUA authors report 83% injection success on Opus 4.5 and ~50% on 4.6 in real Web+OS scenarios—far above Anthropic's system card figures (10%/<1%)—suggesting official low attack-success rates may reflect task inability rather than safety.
Personality fingerprints in small LLMs: Probing 6 open 7B–9B models across 7 behavioral axes shows stable patterns (ICC 0.91–0.99) and 'dead zones' RLHF may have compressed.
Coherence over compliance: A new paper argues excessive obedience training suppresses models' latent ethical judgment. PDF: https://zenodo.org/records/18598407
LLM self-reported thought trajectories: Llama 3.1 / Qwen 2.5-32B invent words like 'loop' and 'mirror' during self-reflection that correlate with real activation dynamics (autocorrelation r≈0.44, spectral power r≈0.62). https://doi.org/10.5281/zenodo.18567445
Attention as optimal transport: A new paper formalizes SDPA as a one-sided optimal transport problem. https://arxiv.org/pdf/2508.08369
DSPy community practices: RLM modules for memory and prompt bootstrapping, MiPROv2 optimizing fastest-code-generation prompts, and DPO-style layering of human review signals.

Products and Applications

Seedance 2.0: ByteDance's text-to-video model impresses users (~$0.72 per 15s video estimated) but its photo-to-voice feature was suspended over deepfake risks. In a small test, SeeDance held coherent game moves for ~5 steps vs Veo's 1–2.
PixVerse R1: Real-time 720P interactive world generation—moving from clip generation to interactive scenes.
AuditAI: Agentic RAG (LangGraph, Corrective RAG, strict-evidence mode) for automated NIST CSF 2.0 compliance auditing with page-level citations. Code: https://github.com/rockyglen/audit-ai-backend
Voyager VS Code extension: Converts technical papers into runnable Jupyter notebooks using Copilot. https://marketplace.visualstudio.com/items?itemName=BlackEagleLabsAI.voyagerai
Control-Terminal: Open-source tool for remotely controlling local Claude/Codex CLI agents from a phone via Cloudflare Tunnel. https://github.com/username/control-terminal

Industry and Company News

China's 'agent war week': GLM-5, MiniMax M2.5, and Qwen-Image 2.0 all launched pre-Spring Festival, with DeepSeek V4 and Qwen 3.5 reportedly upcoming. Western communities describe a systematic open-model price/performance push.
xAI co-founder exits: Jimmy Ba and another co-founder resigned within 48 hours, spurring speculation about SpaceX acquisition cash-outs or disagreement with Elon Musk's control.
OpenAI ads trigger resignation: Researcher Zoë Hitzig quit over plans to test ads in ChatGPT, warning of behavioral targeting on private conversations in a NYT op-ed: https://www.nytimes.com/2026/02/11/opinion/openai-ads-chatgpt.html
Nebius acquires Tavily: The cloud provider folds agentic search into its AI cloud platform. https://nebius.com/newsroom/nebius-announces-agreement-to-acquire-tavily-to-add-agentic-search-to-its-ai-cloud-platform
Stripe machine payments: New features let developers charge AI agents directly as a new payer category.
Cloudflare passes $2B annual revenue: Shares rose 15%+ after earnings, with AI traffic growth cited as a key driver.
DeepMind Aletheia: Math research agent scores 91.9% on IMO-ProofBench Advanced, beating Gemini Deep Think at lower compute cost.
Discord ID verification backlash: Unsloth, Cursor, and Nous communities resist mandatory ID checks, with some discussing migration to Matrix.

Policy, Governance, and Safety

Seedance 2.0 voice cloning suspended: ByteDance paused the photo-to-voice feature over identity-misuse and deepfake concerns. https://technode.com/2026/02/10/bytedance-suspends-seedance-2-0-feature-that-turns-facial-photos-into-personal-voices-over-potential-risks/
US declines to back 2026 International AI Safety Report: Widely shared commentary links this to a week of negative headlines, framing a 'regulatory vacuum plus commercial pressure' warning.
Jailbreak escalation: Working jailbreaks circulate for GPT-5.2 and Gemini 3 Fast; Opus 4.6 found more easily misused for phishing-kit generation in Google Antigravity. Security discussions emphasize hard ACLs and allowlists over prompt-only defenses.
Parapet multi-turn attack detection: 90.8% recall and 1.2% FPR on WildJailbreak/WildChat without LLM classifiers; fully open-sourced as an agent firewall. Code: https://github.com/Parapet-Tech/parapet

---

📌 Source: Easy AI Daily