Easy AI Daily Digest | February 12, 2026: GLM-5 Release, China's Agent War Week, and Agentic Tooling Breakthroughs

Forum topic · 小凯 · 2026-03-27

Summary

The February 12, 2026 edition of Easy AI Daily covers a packed news cycle led by Z.ai's release of GLM-5, a 744B-parameter (40B active) open-weight MoE model with 28.5T training tokens, ~200K context, DeepSeek Sparse Attention, and MIT licensing, which topped open-model leaderboards though priced well above DeepSeek and Kimi alternatives. The digest highlights China's 'Agent War Week' with MiniMax M2.5, Qwen-Image 2.0 (a 7B unified generation/editing model at native 2K), and DeepSeek's apparent V4 Lite update adding 1M-token context. Tooling news includes MCP support in llama.cpp, Karpathy's agent-driven code extraction demo, mini-SWE-agent 2.0, and OpenAI's 1,500-PR Codex workflow. Infrastructure items cover Unsloth's 12x faster MoE Triton kernels, torchao v0.16.0, and multi-GPU hobbyist builds. Security and policy notes include ByteDance suspending Seedance 2.0's photo-to-voice cloning over deepfake risks, disputes over Anthropic's computer-use safety claims, a $3M open evaluation fund from Snorkel and partners, and jailbreak escalation. Industry moves include xAI co-founder departures, an OpenAI researcher resigning over ads, Nebius acquiring Tavily, Stripe enabling machine payments, and DeepMind's Aletheia math agent scoring 91.9% on IMO-ProofBench Advanced.

Easy AI Daily | February 12, 2026

A daily digest of AI industry developments, translated and structured from the original Chinese forum post.

Key Points

Models & Capabilities

Z.ai releases GLM-5: Upgrades from 355B MoE / 32B active to 744B total / 40B active parameters, pretrained on 28.5T tokens with DeepSeek Sparse Attention, ~200K context, and MIT-licensed open weights (~1.5TB BF16). It tops open models on Text Arena, scores 50 on Artificial Analysis' Intelligence Index, posts GDPVal-AA ELO 1412, and reportedly has the lowest hallucination rate in its evaluations.
Links: Official blog | Hugging Face weights | ModelScope | OpenRouter
Compute bottleneck: Formerly rumored as "Pony Alpha," GLM-5 targets agent and long-task scenarios. Traffic surged 10x and Z.ai publicly admits GPU shortage, prioritizing Coding Plan Pro and adjusting pricing. Community notes GPU scarcity now constrains Z.ai, OpenAI, and Google alike. (Z.ai statement, Reddit discussion)
Pricing: Community estimates ~$0.80/M input tokens (~3x DeepSeek V3.2 Speciale, ~1.8x Kimi K2.5) and ~$2.56/M output (~6x DeepSeek) — positioned as premium quality rather than best value. (Cost comparison thread)
GLM-OCR: Open-source OCR model that outperforms Gemini 3 Flash on some tasks in community tests; a cheaper self-hostable option for document/invoice products. (Model card)
DeepSeek "V4 Lite" + 1M context: Reports of a chat-side update to 1M-token context with knowledge cutoff of May 2025, text-only. Users report strong long-document performance but latency rising from ~30s to ~160s on some queries. (Discussion)
DeepSeek's open-source dominance: Fine-grained sparse MoE, MLA, practical sparse attention, DeepEP, and the R1/GRPO recipe are now reused by nearly all frontier open models, including GLM-5. (Analysis thread)
MiniMax M2.5 & StepFun Step-Flash-3.5: M2.5 focuses on task decomposition and long-duration execution; Step-Flash-3.5 claims #1 on MathArena. Both seen as strong speed/cost ratios for coding and math at scale.
Qwen-Image 2.0: Alibaba's 7B unified image generation + editing model with native 2K resolution; shrunk from 20B, renders complex English/Chinese typography and multi-panel comic character consistency. (Announcement)
Qwen3-Coder-Next 80B on an NAS iGPU: Ran on TrueNAS SCALE + Ryzen AI iGPU via llama.cpp Q4_K_M with Vulkan offload and flash attention at ~18 tok/s (tip: remove --no-mmap to fully map into shared memory). (Build details)
Unsloth Triton MoE kernels: Claims up to 12x faster MoE training with ~35% less memory without precision loss, from consumer (3090) to datacenter GPUs, using the new torch._grouped_mm interface. (Details)

Agents & Tooling

Karpathy's new paradigm: Used DeepWiki MCP + GitHub CLI to have 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 MCP servers (GitHub, HF, Exa Search, etc.) via HTTP/WebSocket, with configurable agent loop rounds; llama-server integration planned. (Testing thread)
noapi-google-search-mcp: Headless Chromium wraps Google Search as an MCP tool — no API key needed; supports image search, reverse image search, Lens, flights, stocks, weather, and more. (Repo)
mini-SWE-agent 2.0: ~100-line agent/environment/model harness for coding-agent benchmarks and RL experiments — a shift toward auditable small frameworks. (Author thread)
OpenClaw: Running tmux, git worktree, and Claude Code entirely through Discord, with context/reflection auto-saved to markdown.
OpenAI Codex engineering: DevRel shares a case of merging ~1,500 PRs with no hand-written code, plus best practices for multi-hour agent workflows — emphasizing harness engineering over new models. (Case study)

Infrastructure & Hardware

$5,000 local LLM builds (2026): Community options include dual 128GB Ryzen AI Max+ boxes for 4-bit inference + QAT LoRA, 4×3090 or 7×AMD V620, and quiet Strix Halo builds. Consensus: 100B-class models with long context need 40–48GB VRAM/unified memory minimum. (Thread)
Multi-GPU hobbyist rigs: 6×3090 (144GB) for 10B-class diffusion training; 8×3090 via x16 bifurcation; 4×4090 on X670 at degraded 2.5GT/s x8/x4 bandwidth — workable but requires dealing with bandwidth and power stability issues.
CuTeDSL vs Triton on Blackwell: GPU MODE Kernelbot data shows CUDA and CuTeDSL submissions dominate; contributors find Triton hard to adapt to MXFP8/NVFP4 formats on Blackwell. (Kernelbot dataset)
torchao v0.16.0: Adds MXFP8 MoE training components with Expert Parallel support, cleans up legacy quantization options, and advances ABI stability. (Release notes)
FlashInfer AI kernel contest: Accepts fully agent-generated and agent-assisted CUDA/CuTeDSL kernels; community debates reward hacking and whether techniques like Programmatic Dependent Launch count as violations.

Research & Methods

$3M open evaluation fund: Snorkel, Hugging Face, Together, Prime Intellect, Factory, Harbor, and PyTorch fund open benchmarks to close the gap between closed internal evals and public leaderboards. (Announcement)
Opus computer-use safety disputed: RedTeamCUA authors report 83% prompt-injection success on Opus 4.5 and ~50% on 4.6 in real Web+OS scenarios — far above Anthropic's system-card figures of 10%/<1% — arguing the low official ASR may reflect capability limits, not genuine safety. (Critique thread)
Small-model "personality fingerprints": Probing 6 open 7B–9B models across 7 behavioral axes shows stable patterns (test-retest ICC 0.91–0.99) and "dead zones" RLHF seemingly collapses. (Discussion)
Coherence over compliance (PDF): Argues over-training compliance suppresses models' latent ethical coherence, potentially making them less safe on complex moral questions.
Self-described thought trajectories: Llama 3.1 / Qwen 2.5-32B invent words like "loop" and "mirror" during extended self-reflection that correlate with real activation dynamics (autocorrelation r≈0.44, spectral power r≈0.62). (Paper)
Attention as optimal transport: A new paper reformulates scaled dot-product attention (SDPA) as a one-sided optimal transport problem. (arXiv)
DSPy community: RLM modules for memory systems, MiPROv2 for optimizing fastest-executing codegen prompts, and DPO-style blending of human review signals — treating prompt engineering as a learnable optimization problem. (RLM example)

Products & Applications

Seedance 2.0: ByteDance's text-to-video model impresses users (past the "uncanny valley"; ~$0.72 per 15s clip estimated), but its photo-to-voice feature was suspended over deepfake risks. (TechNode report)
SeeDance vs Veo consistency test: In a tic-tac-toe video test, Veo maintained coherent moves for 1–2 steps; SeeDance held ~5 steps — a useful temporal-consistency benchmark idea. (Test)
PixVerse R1: Real-time 720P interactive world generation — a shift from offline clips to interactive scenes. (Announcement)
AuditAI: LangGraph-based Agentic RAG (Corrective RAG, strict evidence policy, page-level citations) auditing security policies against NIST CSF 2.0. (Code)
Voyager VS Code extension: Converts technical papers into runnable Jupyter notebooks with Copilot. (Marketplace)
Control-Terminal: Remotely control local Claude/Codex CLI agents from a phone via persistent sessions and Cloudflare Tunnel. (Repo)

Industry & Company News

China's "Agent War Week": GLM-5, MiniMax M2.5, and Qwen-Image 2.0 all shipped before Spring Festival, with DeepSeek V4 and Qwen 3.5 reportedly ahead — overseas communities describe it as a coordinated performance-and-price assault. (Reddit thread)
xAI co-founders depart: Jimmy Ba and another co-founder left within 48 hours; speculation ranges from post-acquisition cash-out to disagreement over Musk's control.
OpenAI ads backlash: Researcher Zoë Hitzig resigns via NYT op-ed over plans to test ads in ChatGPT, warning of behavioral targeting on private conversations. (Op-ed)
Nebius acquires Tavily: The cloud provider adds "Agentic Search" to its AI cloud platform. (Press release)
Stripe machine payments: New functionality lets developers charge AI agents directly as a new payer category.
Cloudflare tops $2B annual revenue: 2025 results beat expectations; stock up 15%+ after hours, with AI traffic cited as a growth driver. (Press release)
DeepMind Aletheia: Math research agent scores 91.9% on IMO-ProofBench Advanced, beating January's Gemini Deep Think at lower compute; extension to physics and CS research planned. (Discussion)
Discord ID verification backlash: Unsloth, Cursor, and Nous communities push back against mandatory identity checks; some consider migrating to Matrix.

Policy, Governance & Safety

Seedance 2.0 voice cloning suspended: The dual-branch diffusion feature could synthesize highly realistic voices from a single facial photo, raising identity-fraud and deepfake concerns.
US declines support for 2026 international AI safety report: Viral roundups pairing this with Anthropic's safety lead departing, xAI exits, and Yoshua Bengio's warnings frame a "regulatory vacuum + commercial pressure" narrative.
Jailbreak escalation: Working jailbreaks circulate for GPT-5.2 and Gemini 3 Fast; Opus 4.6 found easier to abuse for phishing content in Google Antigravity. Community emphasizes hard ACLs/allowlists over prompt-based defenses.
Parapet multi-turn attack detection: A scoring formula achieves 90.8% recall at 1.20% FPR on WildJailbreak and WildChat without LLM-based classifiers; fully open-sourced as a front-line firewall for agent systems. (Paper, Code)

---

📌 Source: Easy AI Daily