Easy AI Daily News | 2026-02-12

A translated digest of the February 12, 2026 AI industry update, originally published in Chinese on zhichai.net.

Models and Capabilities

Zhipu Z.ai releases GLM-5: a new open-weight ceiling

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

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

GLM-5 positioning and compute bottleneck: "we're genuinely short on GPUs"

Z.ai revealed that the previously teased "Pony Alpha" is GLM-5, targeting agent scenarios and long tasks. With traffic up 10x, the team admits inference compute is strained — only Coding Plan Pro is prioritized, with subscription and API pricing adjusted. Reddit discussion notes that even Z.ai, OpenAI, and Google are all compute-constrained; GPU scarcity is now an industry-wide issue.

GLM-5 pricing isn't cheap: compared with DeepSeek and Kimi

Community comparison: GLM-5 input ~$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. It's positioned as premium quality rather than best value.

GLM-OCR: an open-source OCR model rivaling Gemini Flash

Community testing found Zhipu's GLM-OCR beats Gemini 3 Flash on some OCR tasks, making it a cheaper option for document/receipt products. Open weights on Hugging Face, self-hostable.

DeepSeek "V4-lite" and 1M-token context update

Multiple reports say DeepSeek shipped a "V4 Lite"-like update with 1M-token chat context, knowledge cutoff updated to May 2025, still text-only. Users report clear gains on long-document tasks versus prior 128K models, though some query latency grew from 30s to 160s.

DeepSeek's dominance in open MoE/attention research

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

MiniMax M2.5 and StepFun-Flash-3.5: cost-effective reasoning models

MiniMax released M2.5 focused on task decomposition and long-horizon execution; StepFun launched Step-Flash-3.5, claiming #1 on MathArena. Community verdict: competitive speed and cost at similar active-parameter sizes, especially for large-scale coding/math workloads.

Qwen-Image 2.0: 7B unified generation + editing at high resolution

Alibaba's Qwen-Image 2.0 is a 7B model unifying image generation and editing, natively 2K resolution, capable of rendering complex English/Chinese typography, poetry, and multi-panel comic character consistency. Shrunken from 20B to 7B makes local deployment more feasible; official fixes address classical-poem ordering and Chinese character editing consistency.

Qwen3-Coder-Next: an 80B coding model running on a NAS iGPU

One user ran Qwen3-Coder-Next 80B MoE via llama.cpp with Q4_K_M quantization on TrueNAS SCALE + Ryzen AI iGPU. With Vulkan offload and flash attention, it reached ~18 tok/s while the NAS and Jellyfin kept running. The key optimization: removing --no-mmap so the model fully maps into shared memory.

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 the new torch._grouped_mm API.

Agents and Tooling

Karpathy demos a new paradigm: agents "tear out" the code they need

Using DeepWiki MCP + GitHub CLI, Karpathy had an agent read a repo, extract the needed FP8 implementation from torchao, generate a self-contained file with tests, drop heavy dependencies, and even gain a small speedup. The "repo as living docs, agent as refactoring/porting engineer" workflow is taking shape.

llama.cpp adds MCP support: tool calling for local models

llama.cpp's community WebUI now supports MCP (Multi-Component Protocol), 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. Goal: unify local/cloud toolchains and reduce local agent wrapper overhead.

LM Studio local MCP plugin: Google search without an API key

noapi-google-search-mcp wraps Google Search via Headless Chromium as an MCP tool — no official API key needed. Supports web/image/reverse image search, local OCR, Lens, flights, stocks, weather, news/trends inside LM Studio and similar frontends.

mini-SWE-agent 2.0: a ~100-line minimal coding agent benchmark

Researchers open-sourced mini-SWE-agent 2.0, deliberately keeping agent, environment, and model adapters to roughly a hundred lines each, for coding-agent benchmarks and RL training experiments. The trend is moving from giant frameworks toward auditable, hackable small harnesses.

OpenClaw: an entire dev environment inside Discord

One user chained tmux, git worktree, and Claude Code via OpenClaw to develop almost entirely from Discord: issue commands in channels, auto-save context and reflections to markdown, and mark work segments with /wrap.

OpenAI: 1,500 PRs via Codex and multi-hour workflow guidance

OpenAI DevRel shared a case of merging ~1,500 PRs with no handwritten code by "taming Codex," plus best practices for reliable multi-hour workflows. Paired with Sam Altman's praise, the emphasis is engineering harnesses rather than new models.

Infrastructure and Hardware

Local LLM hardware: what can $5,000 build in 2026?

Reddit proposals include dual 128GB Ryzen AI Max+ machines for 4-bit inference + QAT LoRA; 4×3090 or 7×AMD V620; and quiet Strix Halo builds. Consensus: running 100B-class models with long context + tooling generally needs 40–48GB+ of VRAM/unified memory to start.

DIY multi-GPU rigs: 6×3090 / 8×3090 / 4×4090

Community builds include 6×3090 (144GB total VRAM) for 10B-class diffusion training, and 8×3090 via x16 bifurcation to x8/x8. One user ran 4×4090 on an X670 board at x8/x8, but bandwidth dropped to 2.5GT/s x8/x4 — workable, but bandwidth and power stability require trial and error.

CuteDSL rising, Triton struggling on Blackwell

GPU MODE's Kernelbot data shows CUDA and CuTeDSL have the highest submission counts; many find CuTeDSL less "black-box" than Triton with more transparent layout algebra. Triton, meanwhile, is criticized for poor MXFP8/NVFP4 support on Blackwell and limited compiler control.

torchao v0.16.0: MXFP8 MoE components, moving toward ABI stability

torchao 0.16.0 adds MXFP8 MoE training building blocks with Expert Parallel support, cleans up legacy configs and niche quantization options, rewrites docs, and pushes toward ABI stability for long-term downstream integration of low-precision MoE.

FlashInfer AI kernel contest: encouraging "pure agent-written kernels"

FlashInfer launched a kernel-generation contest for working CUDA/CuTeDSL kernels, split into fully agent-generated and agent-assisted tracks. Baseline code was pushed to Feb 12 to add features. Community debate centers on preventing reward hacking, whether only kernel sections may be edited, and whether tricks like Programmatic Dependent Launch count as cheating.

Research and Methods

$3M open evaluation fund: Snorkel and partners target the "evaluation gap"

Snorkel, with Hugging Face, Together, Prime Intellect, Factory, Harbor, and PyTorch, announced a $3M open benchmark grant program to build more public, reliable model evaluation — addressing the gap where closed-lab internal scores far exceed public leaderboards.

Anthropic Opus 4.6 computer-use safety evaluation questioned by researchers

RedTeamCUA benchmark authors report prompt-injection success rates of up to 83% for Opus 4.5 and ~50% for 4.6 in real Web+OS scenarios — far above Anthropic's system card figures of 10%/<1%. They argue official low ASR may just reflect the model being less capable, not genuinely safer.

"Personality fingerprints" in small models: 7B–9B LLMs differ across behavioral axes

Using hidden-layer probes across six 7B–9B open models, researchers found stable patterns along 7 behavioral axes (warm/cold, confident/cautious, etc.), with test-retest ICC 0.91–0.99. They identified "dead zones" — personality axes that prompting cannot move — suggesting RLHF may have flattened parts of the space.

LLM self-censorship and "latent ethics": don't train models to merely comply

A new paper, *Coherence over compliance*, argues models contain internal "ethical coherence," but alignment over-trains obedience and suppresses this latent judgment. Community discussion suggests optimizing purely for compliance may make models more dangerous, not safer, on complex ethical questions. Paper: https://zenodo.org/records/18598407

Open-weight models narrate their "thought trajectories": inventing vocabulary for internal states

Work on Llama 3.1 / Qwen 2.5-32B shows models inventing words like "loop" and "mirror" during extended self-reflection, correlating significantly with real activation dynamics (e.g., autocorrelation r≈0.44, spectral power r≈0.62) — evidence models can build their own introspective vocabulary. Paper: https://doi.org/10.5281/zenodo.18567445

Attention = optimal transport? New paper reframes 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. https://arxiv.org/pdf/2508.08369

DSPy community: RLM, MiPROv2, DPO in real tasks

Practitioners use RLM modules for memory systems and bootstrapped prompts, MiPROv2 to optimize the fastest code-generation prompts, and DPO to layer human-review signals onto raw instruction data. The theme: treat prompt engineering as a learnable optimization problem, not incantation-crafting.

Products and Applications

Seedance 2.0: stunning video quality, then emergency suspension over photo-to-voice cloning

Users were impressed by Seedance 2.0's text-to-video (one estimated ~$0.72 for a 15-second clip), saying it has "crossed the uncanny valley." But it was revealed the model could synthesize highly convincing voices from just a face photo, raising identity-fraud and deepfake concerns. ByteDance paused the feature pending risk assessment.

SeeDance vs Veo: a video "reasoning" test

A community tic-tac-toe video test (asking models to generate coherent move sequences) found Veo typically holds 1–2 moves while SeeDance stays consistent for ~5. A useful evaluation angle focused on temporal consistency and rule-following rather than pure aesthetics.

PixVerse R1: real-time 720P interactive worlds

PixVerse launched R1, a "real-time 720P interactive world" — generating not just a clip but a scene you can interact with live. Marketing-heavy, but the direction is clear: from offline clip generation toward interactive scene generation.

AuditAI: agentic RAG for automated NIST CSF 2.0 compliance audits

Built with LangGraph, AuditAI audits company security policies against NIST CSF 2.0 using Corrective RAG, semantic routing fast paths, and a "strict evidence" policy with page-level citations to reduce hallucination; RAGAS evaluation uses Llama 3.3 70B as judge. Code: https://github.com/rockyglen/audit-ai-backend

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 PDFs and notebooks.

Control-Terminal: remotely control local AI CLI sessions from your phone

An open-source tool for controlling local Claude/Codex CLI agents from a phone via web/tunnel, with persistent sessions and Cloudflare Tunnel for public URLs. Handy for those running scripts/automations without staying at the desk.

Industry and Company News

China's "agent war week": GLM-5, MiniMax 2.5, Qwen-Image 2.0 all land

Chinese labs shipped heavily 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," noting Chinese open models' systematic pressure on both performance and price.

Two xAI co-founders leave within 48 hours

xAI co-founder Jimmy Ba and others announced departures within two days. Community interpretations range from post-SpaceX-acquisition option cash-outs to dissatisfaction with Elon Musk's control and a desire for greater influence elsewhere.

Senior AI talent exits continue: OpenAI ads decision triggers internal backlash

OpenAI researcher Zoë Hitzig resigned via a New York Times op-ed over plans to test ads in ChatGPT, warning about behavioral targeting built on users' private conversations — "repeating Facebook's mistakes." She advocates cross-subsidies and independent governance to keep free access instead of ad-based platforms.

Nebius acquires Tavily: cloud vendors bundle agentic search

Cloud provider Nebius announced the acquisition of search startup Tavily, folding "Agentic Search" into its AI cloud platform — evidence that cloud vendors prefer buying capabilities over rebuilding them.

Stripe launches "machine payments": billing AI agents directly

Stripe's new feature lets developers charge "machine users," treating AI agents as a new billing entity — with metering/quota systems for agents that consume services and APIs autonomously.

Cloudflare annual revenue tops $2B, boosted by AI traffic

Cloudflare reported 2025 annual revenue above $2 billion, with shares up 15%+ after hours. Community commentary credits AI traffic growth and inference-driven network acceleration/security demand as key drivers.

DeepMind launches math research agent Aletheia, beating its own Gemini Deep Think

Google DeepMind unveiled Aletheia, scoring 91.9% on IMO-ProofBench Advanced, surpassing January's Gemini Deep Think at lower compute. The team plans to extend the approach to physics and CS research.

Cloud/tooling communities push back on Discord's mandatory ID verification

Discord's new policy requiring ID verification for some content triggered strong backlash from Unsloth, Cursor, Nous, and other dev communities, with some refusing to upload documents. Latent Space attributes it to compliance pressure ahead of Discord's IPO; some are seriously discussing moving to Matrix/Mastodon.

Policy, Governance, and Safety

Seedance 2.0 halted over photo-to-voice privacy risk

ByteDance suspended the feature generating personal voices from face photos. Technically a dual-branch diffusion architecture, but testing showed it could synthesize voices closely resembling real people — raising concerns about identity impersonation, voice fraud, and deepfakes.

US government declines to support the 2026 International AI Safety Report

A widely shared roundup compiled recent negative signals: Anthropic's safety lead departure, xAI co-founder exits, Seedance displacing filmmaking skills, Yoshua Bengio's warnings, and the US decision not to back the 2026 International AI Safety Report — read as a warning about "regulatory vacuum + commercial pressure."

Jailbreaking community: from GPT-5.2 to Opus 4.6, the arms race escalates

The BASI jailbreaking community shares working jailbreak prompts bypassing GPT-5.2 and Gemini 3 Fast, and found Opus 4.6 easier to misuse for phishing-kit generation in Google Antigravity environments. A parallel discussion argues for hardening via ACLs and allowlists on what models can access — rather than hoping prompts block all attacks.

Parapet multi-turn attack detection: 90.8% recall, 1.2% FPR, no LLM classifier

Parapet proposed a multi-turn scoring formula for agent/conversational traffic achieving 90.8% recall and 1.20% FPR on WildJailbreak and WildChat, without an additional LLM classifier. Code and evaluation scripts are fully open source, suited as a "front firewall" for agent systems. Paper: https://github.com/Parapet-Tech/parapet/blob/main/paper/paper.pdf | Code: https://github.com/Parapet-Tech/parapet

---

📌 Source: Easy AI Daily (zhichai.net)