Easy AI Daily News Digest | February 12, 2026: GLM-5 Launch, China's Agent War Week, and More

Models & Capabilities

Zhipu Z.ai Releases GLM-5: A New Open-Weight Ceiling

GLM-5 upgrades from 355B MoE/32B active to 744B/40B active parameters, with pretraining data expanded to 28.5T tokens. It introduces DeepSeek Sparse Attention, supports ~200K context with ultra-long outputs, and is open-sourced under MIT license (BF16 weights ~1.5TB). It tops Text Arena among open models, scores 50 on the Artificial Analysis Intelligence Index, hits GDPVal-AA ELO 1412 on white-collar benchmarks, and has the lowest hallucination rate in their evaluations.

• Official blog | Hugging Face weights | ModelScope weights | OpenRouter | Artificial Analysis | vLLM support

GLM-5 Positioning and Compute Bottleneck: "We Really Don't Have Enough GPUs"

Z.ai revealed that the previously teased "Pony Alpha" is GLM-5, targeting agent scenarios and long tasks. With traffic up 10x, the company stressed inference compute constraints, prioritizing Coding Plan Pro access and adjusting subscription/API pricing. Reddit discussion notes that GPU scarcity affects Z.ai, OpenAI, and Google alike.
• Z.ai compute/billing note | Reddit thread

GLM-5 Pricing: Not Cheap vs DeepSeek and Kimi

Community comparison: GLM-5 input ~$0.80/M tokens — roughly 3x DeepSeek V3.2 Speciale and 1.8x Kimi K2.5; output ~$2.56/M tokens — about 6x DeepSeek and slightly pricier than Kimi. Positioned for premium quality rather than best value.
• Reddit cost discussion

GLM-OCR: Open OCR Model Rivaling Gemini Flash

Community testing found Zhipu's GLM-OCR beats Gemini 3 Flash on some OCR tasks, viewed as a cheaper option for document/receipt products. Open weights on Hugging Face, self-hostable.
• Model card | Latent Space feedback

DeepSeek "V4-lite" and 1M-Context Update

Multiple reports say DeepSeek shipped a "V4 Lite"-like update with 1M-token chat context and knowledge cutoff of May 2025, still text-only. Users report clear gains on long-document tasks, though some query latency stretched from 30s to 160s.
• Reddit thread

DeepSeek's Dominance in Open MoE/Attention Research

Developers catalogued DeepSeek's technical output: fine-grained sparse MoE, MLA, practical sparse attention, DeepEP inference library, and the public R1/GRPO recipe. Nearly all frontier open models (including GLM-5) now reuse these components.
• Influence thread

MiniMax M2.5 and StepFun-Flash-3.5: Cost-Effective Reasoning Models

MiniMax released M2.5 focused on task decomposition and long execution; StepFun launched Step-Flash-3.5, claiming #1 on MathArena. Community verdict: competitive speed and cost at similar active parameters, suited for high-volume coding/math workloads.
• MiniMax | StepFun

Qwen-Image 2.0: 7B Unified Generation + Editing at 2K Native

Alibaba's Qwen-Image 2.0 is a 7B unified gen/edit model with native 2K resolution, rendering complex English/Chinese typography, poetry, and multi-panel comic character consistency. Shrunken from 20B to 7B, more feasible for local runs; ordering and Chinese-text editing issues reportedly fixed.
• Official tweet | Reddit

Qwen3-Coder-Next: 80B Code Model on a NAS iGPU

A user ran Qwen3-Coder-Next 80B MoE via llama.cpp with Q4_K_M quantization on TrueNAS SCALE + Ryzen AI iGPU, reaching ~18 tok/s with Vulkan offload and flash attention while NAS and Jellyfin ran concurrently. Key tweak: removing --no-mmap to fully map the model into shared memory.
• Reddit details

Unsloth Triton MoE Kernels: 12x Faster Training, 30%+ Less Memory

Unsloth's MoE Triton kernels claim up to 12x MoE training speedup and ~35% memory reduction without precision loss, from consumer (3090) to datacenter GPUs, using PyTorch's new torch._grouped_mm interface.
• Reddit | Official tweet

Agents & Tooling

Karpathy Demos Agents That "Rip Out" the Code They Need

Using DeepWiki MCP + GitHub CLI, Karpathy had an agent read a repo, extract a needed torchao FP8 implementation into a self-contained file with tests, drop heavy dependencies, and gain a small speedup — a "repo as ground truth, agent as refactoring/porting worker" workflow.
• Tweet

llama.cpp Adds MCP Support

The community llama.cpp WebUI now supports MCP with configurable agent loop rounds and tool preview lines, calling remote MCP servers (GitHub, HF, Exa Search) over HTTP/WebSocket; llama-server integration is planned to unify local/cloud tooling.
• Reddit

noapi-google-search-mcp: Keyless Google Search for LM Studio

Wraps Google Search via headless Chromium as an MCP tool — no API key needed — exposing web, image, reverse image, local OCR, Lens, flights, stocks, weather, and news/trends to local models.
• Repo

mini-SWE-agent 2.0: ~100-Line Coding Agent Harness

Released for coding-agent benchmarking and RL experiments, keeping agent, environment, and model integration at ~100 lines each — a trend toward auditable, hackable harnesses over giant frameworks.
• Author tweet

OpenClaw: Full Dev Environment Inside Discord

Chains tmux, git worktree, and Claude Code so the developer works almost entirely from Discord: issue commands in channels, auto-save context/reflections to markdown, mark work sessions with /wrap.
• Discussion

OpenAI: 1,500 Codex PRs and Multi-Hour Workflow Guide

OpenAI DevRel shared merging ~1,500 PRs with no hand-written code via a tamed Codex, plus best practices for reliable multi-hour workflows — emphasizing engineering harnesses over new models.
• Case study | Guide

Infrastructure & Hardware

What $5,000 Buys for Local LLMs in 2026

Options discussed: dual 128GB Ryzen AI Max+ machines for 4-bit inference + QAT LoRA; 4×3090 or 7×AMD V620; quiet Strix Halo builds. Consensus: 100B-class, long-context + tooling setups need 40–48GB VRAM/unified memory minimum.
• Thread

DIY Multi-GPU Builds: 6×3090, 8×3090, 4×4090

Examples: 6×3090 (144GB total) for 10B-class diffusion training; 8×3090 via x16 bifurcation; 4×4090 on X670 with x8/x8 at reduced 2.5GT/s — workable but with bandwidth and power-stability pitfalls.
• 6×3090 build

CuteDSL Rising; Triton Struggles on Blackwell

GPU MODE Kernelbot data shows CUDA and CuTeDSL submissions lead; contributors find CuTeDSL less black-box with more transparent layout algebra. Triton is criticized for poor Blackwell MXFP8/NVFP4 support and limited compiler control.
• Kernelbot dataset

torchao v0.16.0: MXFP8 MoE Blocks, Toward ABI Stability

Adds MXFP8 MoE training building blocks with Expert Parallel support, cleans up legacy configs, rewrites docs, and advances ABI stability for downstream low-precision MoE integration.
• Release notes

FlashInfer AI Kernel Contest Encourages "Pure Agent-Written Kernels"

Submissions of working CUDA/CuTeDSL kernels in fully-agent-generated and agent-assisted tracks; community debate covers reward cheating, kernel-only edits, and whether Programmatic Dependent Launch counts as fair play.

Research & Methods

$3M Open Evaluation Fund

Snorkel, with HF, Together, Prime Intellect, Factory, Harbor, and PyTorch, announced a $3M open benchmark grant program to close the gap between closed-lab internal scores and public leaderboards.
• Announcement

Anthropic Opus 4.6 Computer-Use Safety Findings Contested

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%), arguing low official ASR may reflect model incapability rather than safety.
• Thread

"Personality Fingerprints" of 7B–9B Models

Probing 6 open models across 7 behavioral axes showed stable patterns (ICC 0.91–0.99) and "dead zones" some axes can't be prompted out of — suggesting RLHF may flatten parts of the space.
• Reddit

LLM Self-Censorship and "Latent Ethics"

The paper "Coherence over compliance" argues overtraining obedience suppresses models' latent ethical coherence, potentially making them less safe on complex moral questions.
• Paper

Open-Weight Models Invent Vocabulary for Internal States

Work on Llama 3.1 / Qwen 2.5-32B shows models coin words like "loop" and "mirror" during extended self-reflection, correlating with real activation dynamics (autocorrelation r≈0.44, spectral power r≈0.62).
• Paper

Attention as Optimal Transport

A new paper formalizes standard dot-product attention (SDPA) as a one-sided optimal transport problem — mathematically elegant, still theoretical.
• arXiv

DSPy Community: RLM, MiPROv2, DPO in Practice

Members use RLM modules for memory systems, MiPROv2 to optimize fastest-codegen prompts, and DPO to layer human-review signals onto instruction data — treating prompting as a learnable optimization problem.
• RLM example

Products & Applications

Seedance 2.0: Stunning Video, Suspended Voice Cloning

ByteDance's Seedance 2.0 impressed on text-to-video (one user computed ~$0.72 per 15-second clip), but its photo-to-voice feature could clone a person's voice from a facial photo; ByteDance suspended it pending risk review.
• TechNode

SeeDance vs Veo: Temporal Consistency Test

A "three-move chess video" test found Veo holds coherent moves for 1–2 steps while SeeDance sustains ~5 — a useful evaluation angle beyond raw aesthetics.
• Test

PixVerse R1: Real-Time 720P Interactive Worlds

PixVerse's R1 shifts from offline clip generation to real-time interactive scene generation — heavy marketing, clear direction.
• Announcement

AuditAI: Agentic RAG for NIST CSF 2.0 Audits

A LangGraph-based system using Corrective RAG, semantic routing, and strict-evidence policies with page-level citations; RAGAS evaluation uses Llama 3.3 70B as judge.
• Code

Voyager VS Code Extension: Papers to Jupyter Notebooks

Uses Copilot to convert technical papers into runnable notebooks with code snippets and insertable custom cells.
• Extension

Control-Terminal: Manage Local AI CLI Agents from Your Phone

Open-source tool for remotely controlling local Claude/Codex CLI sessions via web/tunnel (Cloudflare Tunnel), with persistent sessions.
• Repo

Industry & Business

China's "Agent War Week": GLM-5, MiniMax 2.5, Qwen-Image 2.0

Chinese labs shipped a wave of releases pre-Spring Festival: Z.ai's GLM-5, MiniMax M2.5, Alibaba's Qwen-Image 2.0, with DeepSeek V4 and Qwen 3.5 reportedly upcoming. Overseas observers describe it as a coordinated performance/price push by Chinese open models.
• Reddit

Two xAI Co-Founders Leave Within 48 Hours

Co-founder Jimmy Ba and others announced departures in two days; speculation ranges from post-acquisition cashouts to dissatisfaction with Elon Musk's control.
• Report

OpenAI Researcher Quits Over Ads Plan

Zoë Hitzig resigned via a NYT op-ed, worried ChatGPT ads built on private conversations would repeat Facebook's mistakes; she advocates cross-subsidies and independent governance instead.
• Op-ed

Nebius Acquires Tavily

The cloud provider is buying the agentic-search startup to bundle search capabilities into its AI cloud platform.
• Press release

Stripe Launches Machine Payments

Stripe now lets developers charge AI agents directly as a new class of paying customer, with billing/quota systems for agent economic activity.
• Announcement

Cloudflare Passes $2B Annual Revenue

Fiscal 2025 results beat $2B; stock rose 15%+ after hours, with AI traffic and inference-driven network/security demand cited as growth drivers.
• Press release

DeepMind's Aletheia Math Agent Beats Its Own Gemini Deep Think

Aletheia scored 91.9% on IMO-ProofBench Advanced with less compute; the team plans to extend the approach to physics and CS research.
• Discussion

Discord ID Verification Backlash

Discord's new ID-verification policy for some content drew pushback from Unsloth, Cursor, and Nous communities, with some considering migration to Matrix/Mastodon amid pre-IPO compliance pressure.
• Policy tweet

Policy, Governance & Safety

Seedance 2.0 Voice Cloning Suspended

The dual-branch diffusion architecture behind photo-to-voice could produce highly realistic personal voices, raising identity-theft and deepfake concerns; ByteDance paused the feature.
• TechNode

US Declines to Back 2026 International AI Safety Report

A widely shared roundup paired Anthropic's safety-lead departure, xAI exits, Seedance's disruption of filmmaking skills, Yoshua Bengio's warnings, and the US decision — framed as regulatory vacuum plus commercial pressure.

Jailbreak Escalation from GPT-5.2 to Opus 4.6

Working jailbreaks for GPT-5.2 and Gemini 3 Fast circulate; Opus 4.6 proved easier to misuse for phishing content in Google Antigravity. Parallel discussions push hard ACL/allowlist controls over prompt-based defenses.

Parapet Multi-Turn Attack Detection: 90.8% Recall, 1.2% FPR

Parapet's multi-turn scoring approach hits 90.8% recall at 1.20% FPR on WildJailbreak and WildChat without extra LLM classifiers; code and eval scripts fully open-source, suitable as a front-line firewall for agent systems.
• Paper | Repo

---

📌 Source: Easy AI Daily