Easy AI Daily | February 12, 2026

A digest of AI industry news compiled by zhichai.net, covering model releases, agent tooling, infrastructure, research, products, industry moves, and safety/policy.

Models and Capabilities

Z.ai Releases GLM-5: A New Ceiling for Open Weights

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

• Official blog | Weights on Hugging Face | Weights on ModelScope | OpenRouter page | Artificial Analysis analysis | LMSYS / vLLM support | vLLM integration

GLM-5 Positioning and Compute Bottleneck: "We're Really GPU-Short"

Z.ai revealed that the previously teased "Pony Alpha" is GLM-5, aimed at agent scenarios and long tasks. With traffic up 10x, officials stressed inference compute constraints, prioritizing Coding Plan Pro and adjusting subscription/API pricing. Reddit discussion notes that even Z.ai, OpenAI, and Google are variously constrained by GPU supply — compute shortage is now industry-wide.
• Z.ai compute/billing statement | Reddit: Z.ai openly GPU-starved | Subscription changes discussion

GLM-5 Pricing: Not Cheap vs. DeepSeek and Kimi

Community comparisons: GLM-5 input pricing ~$0.80/M tokens — about 3x DeepSeek V3.2 Speciale and 1.8x Kimi K2.5; output ~$2.56/M tokens — about 6x DeepSeek and slightly above Kimi. Positioned for premium quality rather than best value.
• Reddit cost comparison | Artificial Analysis cost/capability chart

GLM-OCR: Open-Source OCR Challenger to Gemini Flash

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

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

Multiple reports indicate DeepSeek shipped a "V4 Lite"-like update: 1M-token chat context, knowledge cutoff updated to May 2025, still text-only. Users report notably stronger long-document performance vs. prior 128K models, though some query latency stretched from 30s to 160s.
• Long-context experience | Reddit discussion 1 | Reddit discussion 2

DeepSeek's Dominance in Open MoE/Attention Lineage

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

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

MiniMax released M2.5, focused on task decomposition and long-duration execution; StepFun launched Step-Flash-3.5, claiming #1 on MathArena. Community verdict: competitive speed and cost at similar active-parameter counts, well-suited to large-scale coding/math workloads.
• MiniMax announcement | Step-Flash-3.5 intro

Qwen-Image 2.0: 7B Unified Generation + Editing at High Resolution

Alibaba's Qwen-Image 2.0 is a 7B unified generation/editing model with native 2K resolution, capable of rendering complex English/Chinese typography, poetry, and multi-panel comic character consistency. Shrunk from 20B to 7B vs. the previous generation, making local deployment more feasible. The team has fixed classical-poem ordering and Chinese-character editing consistency issues.
• Qwen announcement | Reddit discussion

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

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

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

Unsloth released MoE Triton kernels claiming up to 12x MoE training speedup and ~35% memory reduction without precision loss, supporting consumer to datacenter GPUs (3090 included), using PyTorch's new torch._grouped_mm interface. Valuable for small teams experimenting with MoE on single machines.
• Reddit charts and discussion | Unsloth announcement

Agents and Tooling

Karpathy Demonstrates a New Paradigm: Agents That "Tear Off" the Code They Need

Using DeepWiki MCP + GitHub CLI, Karpathy had an agent read a repo, extract the needed torchao FP8 implementation into a self-contained file with tests, drop heavy dependencies, and even gain a small speedup. The "repo as source of truth, agent as refactoring/porting engineer" workflow is taking shape.
• Karpathy's example

llama.cpp Adds MCP Support: Tool Calling for Local Models

llama.cpp's WebUI now supports MCP (Model Context Protocol), with configurable agent loop counts, tool preview lines, and remote MCP servers over HTTP/WebSocket (GitHub, HF, Exa Search); llama-server integration is planned. Goal: unify local/cloud toolchains and reduce local agent glue costs.
• Reddit release thread

LM Studio / Local MCP Plugin: Google Search Without API Keys

noapi-google-search-mcp wraps Google Search into an MCP tool via headless Chromium — no official API key needed. From LM Studio and similar frontends you can call web search, images, reverse image search, local OCR, Lens, flights, stocks, weather, news/trends, effectively giving local models a browser.
• Repository

mini-SWE-agent 2.0: A ~100-Line Minimal Coding Agent for Benchmarks

Researchers open-sourced mini-SWE-agent 2.0, deliberately keeping agent, environment, and model interface each around a hundred lines, for coding-agent benchmarking and RL experiments. The trend is moving from giant frameworks toward auditable, hackable small harnesses.
• Author's post

OpenClaw: A Full Dev Environment Inside Discord

One developer chained tmux, git worktree, and Claude Code via OpenClaw, working almost entirely in Discord: issuing commands in channels, auto-saving context and reflections to markdown, and marking milestones with /wrap — treating "workflow + auditable context" as first-class.
• OpenClaw discussion

OpenAI: 1,500 PRs with Codex, Plus Multi-Hour Workflow Guidance

OpenAI DevRel shared a case of landing ~1,500 PRs with no hand-written code by "taming" Codex, plus best practices for reliable multi-hour workflows. Alongside Sam Altman's praise of Codex, the emphasis is engineering the harness rather than building a new model.
• Engineering case | Multi-hour workflow tips | Sam Altman comment

Infrastructure and Hardware

Local LLM Hardware: What Can $5,000 Buy in 2026?

Reddit discussion of $5,000 local LLM builds included: dual 128GB Ryzen AI Max+ machines for 4-bit inference + QAT LoRA; 4x3090 or 7x AMD V620; silent Strix Halo options. Consensus: running 100B-class models with long context + tooling generally needs 40–48GB+ VRAM/unified memory to start.
• Build discussion

DIY Multi-GPU Rigs: 6x3090 / 8x3090 / 4x4090 Configurations

Community builds: 6x3090 (144GB total VRAM) for 10B-class diffusion training; 8x3090 via x16 bifurcation into x8/x8; 4x4090 on an X670 board with x8/x8 bifurcation but bandwidth reduced to 2.5GT/s x8/x4. Workable, but bandwidth and power stability require hands-on troubleshooting.
• 6x3090 build | 4x4090 x8/x8 discussion

CuteDSL Rising; Triton Struggles on Blackwell

GPU MODE Kernelbot stats show CUDA and CuTeDSL with the highest submission counts; many find CuTeDSL less of a black box than Triton with more transparent layout algebra. Triton is criticized for poor support of new Blackwell formats like MXFP8/NVFP4 and insufficient compiler control; fixes await upstream improvements.
• Kernelbot dataset | Community discussion

torchao v0.16.0: MXFP8 MoE Building Blocks, ABI Stability Progress

torchao 0.16.0 adds MXFP8 MoE training components with Expert Parallel support, cleans up legacy configs and niche quantization options, rewrites docs/README, and emphasizes progress toward ABI stability for long-term downstream integration of low-precision MoE.
• Release notes

FlashInfer AI Kernel Contest: Encouraging "Pure Agent-Written Kernels"

FlashInfer launched a kernel-generation contest for working CUDA/CuTeDSL kernels, with fully agent-generated and agent-assisted tracks. Baselines were delayed to Feb 12 for feature additions; discussion centers on preventing reward gaming, whether only kernel sections may be modified, and whether advanced tricks like Programmatic Dependent Launch count as cheating.
• Contest rules and discussion

Research and Methods

$3M Open Evaluation Fund: Snorkel et al. Address the "Evaluation Gap"

Snorkel, with HF, Together, Prime Intellect, Factory, Harbor, PyTorch and others, announced a $3M open benchmark grant program to build more public, reliable model evaluation — targeting the gap where closed labs' internal scores far exceed public leaderboards.
• Announcements | 2 | 3

Anthropic Opus 4.6 Computer-Use Safety Eval Questioned

RedTeamCUA benchmark 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 of 10% / <1% — arguing the official low attack-success rates may reflect task incompetence rather than genuine safety.
• Researcher critique thread

Personality Fingerprints in Small Models: 7B–9B LLMs Differ Across Behavioral Axes

Probing 6 open 7B–9B models with hidden-layer probes found stable patterns across 7 behavioral axes (warmth/coldness, confidence/caution, etc.) with test-retest ICC 0.91–0.99. The study found "dead zones" — personality axes unreachable by any prompting — suggesting RLHF may collapse parts of the behavioral space.
• Reddit paper/tool intro

LLM Self-Censorship and "Latent Ethics": Don't Train Models to Only Comply

A new paper, *Coherence over compliance*, argues large models have an internal ethical coherence that over-training for obedience suppresses. HF community discussion notes that optimizing purely for compliance may make models more dangerous on complex ethical questions, not safer.
• Paper PDF

Open-Weight Models Describe Their Own "Thought Trajectories"

Work shared in the Eleuther community on Llama 3.1 / Qwen 2.5-32B: during extended self-reflection, models invent words like "loop" and "mirror" that correlate significantly with real activation dynamics (e.g., autocorrelation r≈0.44, spectral power r≈0.62), indicating models can build their own introspective vocabulary.
• Paper

Attention = Optimal Transport? A New Reformulation of SDPA

A new paper formalizes standard dot-product attention (SDPA) as a one-sided optimal transport problem, offering a fresh theoretical lens for optimizing attention and designing sparse/approximate variants. Community verdict: mathematically elegant, still theoretical.
• arXiv paper

DSPy Community: RLM, MiPROv2, DPO on Real Tasks

DSPy Discord users are applying RLM modules to memory systems and bootstrapped prompts, using MiPROv2 to optimize for the fastest code-generation prompts, and layering DPO-style human review signals onto raw instruction data. The theme: treating prompt engineering as a learnable optimization problem, not hand-crafted incantations.
• DSPy RLM code example | Kaggle/AIMO_v3 thread

Products and Applications

Seedance 2.0: Stunning Video, But Photo-to-Voice Clone Pulled

Users were impressed by Seedance 2.0's text-to-video quality (past the "uncanny valley"; some estimated ~$0.72 per 15-second clip). But after reports it could synthesize highly realistic voices from just a facial photo — raising identity-fraud and deepfake concerns — ByteDance suspended the feature pending risk assessment.
• Quality discussion | TechNode report

SeeDance vs. Veo: A Small "Reasoning" Video Test

A "tic-tac-toe video" test asked models to generate coherent move sequences: Veo typically stayed consistent for 1–2 moves; SeeDance held to ~5. A single test, but a useful lens on temporal consistency and rule adherence over pure aesthetics.
• Comparison test

PixVerse R1: Real-Time 720P Interactive Worlds

PixVerse launched R1, pitching "real-time 720P interactive worlds" — not just generating a clip, but interacting inside it in real time. Marketing-heavy, but the direction is clear: from offline clip generation toward interactive scene generation.
• PixVerse announcement

AuditAI: Agentic RAG for NIST CSF 2.0 Compliance Audits

A community project built "AuditAI" with LangGraph: Agentic RAG auditing company security policies against NIST CSF 2.0, using Corrective RAG, semantic routing fast paths, and a strict-evidence strategy with page-level citations to reduce hallucination. RAGAS evaluation used Llama 3.3 70B as judge.
• Backend code

Voyager VS Code Extension: Turn Papers into Jupyter Notebooks

Voyager uses Copilot in VS Code to convert technical papers into runnable Jupyter notebooks with code snippets and insertable custom cells for section-by-section understanding — for those tired of switching between PDF and notebook.
• Marketplace listing

Control-Terminal: Remotely Drive Local AI CLI Sessions from Your Phone

Control-Terminal is an open-source tool for controlling local Claude/Codex CLI agents from a phone over web/tunnel, with persistent sessions and public URLs via Cloudflare Tunnel — handy for anyone running scripts/automations away from their desk.
• Repository | Docs

Industry and Company News

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

Chinese vendors launched heavily the week before Spring Festival: Z.ai's GLM-5, MiniMax M2.5, Alibaba's Qwen-Image 2.0, with DeepSeek V4 and Qwen 3.5 reportedly on the way. Overseas observers describe the pace as "bloody" — a coordinated performance-and-price push from Chinese open models.
• Reddit roundup | News discussion

Two xAI Co-Founders Leave Within 48 Hours

xAI co-founder Jimmy Ba and others announced departures within two days, triggering speculation: post-SpaceX-acquisition cash-outs, or disagreements over Elon Musk's control and desire for greater influence elsewhere.
• Jimmy Ba departure thread

Senior AI Talent Exodus Continues: OpenAI's Ads Decision Sparks Internal Backlash

OpenAI researcher Zoë Hitzig resigned in a New York Times op-ed over plans to test ads in ChatGPT, warning of behavioral targeting built on users' vast private conversations — repeating Facebook's mistakes. She advocates cross-subsidies and independent governance to keep free access rather than ad-based monetization.
• Hitzig op-ed | Reddit discussion

Nebius Acquires Tavily: Cloud Vendors Bundle Agentic Search

Cloud provider Nebius announced the acquisition of agentic-search startup Tavily, folding the capability into its AI cloud platform. Tavily raised funding just last year; the deal shows cloud vendors prefer buying ready-made capabilities over rebuilding.
• Press release

Stripe Launches "Machine Payments": Billing AI Agents Directly

Stripe's new feature lets developers charge "machine users" directly, treating AI agents as a new payer category. The framing: when agents consume services and APIs on their own, they need billing/quota systems to manage their economic behavior.
• Feature intro

Cloudflare Tops $2B Annual Revenue, Boosted by AI Traffic

Cloudflare reported 2025 annual revenue above $2 billion, with shares up over 15% after hours. Community commentary credits AI-driven traffic growth and inference-related demand for acceleration and security as key growth drivers.
• Earnings press release

DeepMind Launches Aletheia Math Research Agent, Outperforming Its Own Gemini Deep Think

Google DeepMind unveiled Aletheia, a math research agent scoring 91.9% on IMO-ProofBench Advanced, beating January's Gemini Deep Think at lower compute cost. The team plans to extend the approach to physics and computer science research.
• Aletheia results discussion

Dev Communities Push Back on Discord's Mandatory ID Verification

Discord's new ID-verification policy for some content drew strong pushback from Unsloth, Cursor, Nous, and other developer communities, with some refusing to upload documents. Latent Space attributes it to pre-IPO compliance pressure; some are seriously discussing migrating to Matrix/Mastodon.
• Discord policy post

Policy, Governance, and Safety

Seedance 2.0 Suspended Over Photo-to-Voice Privacy Risk

ByteDance paused Seedance 2.0's "generate a personal voice from a facial photo" feature. Technically a dual-branch diffusion architecture, in practice it synthesized voices highly similar to real people, raising regulatory and ethical concerns over identity impersonation, voice fraud, and deepfakes.
• TechNode report | Reddit discussion

US Government Declines to Back the 2026 International AI Safety Report

A widely shared meme collaged recent negative news: Anthropic's safety lead departing, xAI co-founder exits, Seedance replacing filmmaker skills, Yoshua Bengio's warnings, and the US decision not to support the 2026 International AI Safety Report — framed as a warning about a "regulatory vacuum plus commercial pressure."
• Discussion

Jailbreak Community: Arms Race from GPT-5.2 to Opus 4.6

The BASI Jailbreaking community has working jailbreak prompts bypassing GPT-5.2 and Gemini 3 Fast, and found Opus 4.6 easier to misuse for harmful outputs like phishing kits in Google Antigravity. Related discussion stresses hardening what models can access via ACLs and allowlists rather than relying on prompts to block all attacks.
• GPT-5.2 jailbreak discussion | ACL/allowlist security discussion

Parapet Multi-Turn Attack Detection: 90.8% Recall, 1.2% FPR Without LLM Classifiers

Parapet proposes a multi-turn scoring formula for agent/conversation attacks, achieving 90.8% recall and 1.20% FPR on WildJailbreak and WildChat with no additional LLM classifier. Code and evaluation scripts are fully open source — suitable as a front-line firewall for existing agent systems.
• Paper PDF | Code repository

---

📌 Source: Easy AI Daily