Easy AI Daily Digest | February 13, 2026: Gemini 3 Deep Think V2, MiniMax M2.5, GLM-5, and More

Forum topic · 小凯 · 2026-03-27

Summary

This February 13, 2026 AI industry digest covers major model releases and ecosystem updates. Google launched Gemini 3 Deep Think V2, scoring 84.6% on ARC-AGI-2, and unveiled Aletheia, a math-specialized agent scoring 100% on IMO 2024 tasks. OpenAI released GPT-5.3-Codex-Spark with Cerebras, generating over 1000 tokens per second, while MiniMax M2.5 (230B-parameter MoE, 80.2% SWE-Bench Verified) and Zhipu's open-weight GLM-5 target agentic coding at a fraction of competitor costs. DeepSeek previewed V4 with 1M-token context. Infrastructure news includes Mooncake joining the PyTorch ecosystem, torchao v0.16.0, and GPU MODE's $20-30K kernel-generation experiment. Research highlights span Anthropic's Introspection work, rank-1 LoRA matching full RL fine-tuning, and GoodfireAI's RLFR. Business updates cover Anthropic's $380B valuation funding round and Perplexity's Deep Research quota cuts.

Easy AI Daily | February 13, 2026

A digest of AI industry news for February 13, 2026, covering models, agent tooling, infrastructure, research, products, business, and safety.

Models & Capabilities

Google Gemini 3 Deep Think V2: 84.6% on ARC-AGI-2

Now available to paid Gemini app users with early access via Vertex AI/Gemini API. Certified by ARC Prize: 84.6% on ARC-AGI-2, 48.4% on HLE, Codeforces Elo 3455, olympiad-level physics/chemistry and IMO performance. Used for semiconductor materials design, math paper review, and sketch-to-CAD/STL workflows. Per-task cost on ARC tasks reportedly reduced to ~18% of the original.

Google Aletheia: Math-Specialized Proof Agent

DeepMind's math research agent scored 91.9% on IMO-ProofBench Advanced and 100% on IMO 2024 tasks, well above general models like Gemini 3 Pro. Uses a generator-plus-verifier architecture; similar methods planned for physics and computer science.

OpenAI GPT-5.3-Codex-Spark: Ultra-Low-Latency Coding Model

Built with Cerebras, in research preview for ChatGPT Pro users via Codex app/CLI/IDE extensions. Claims 1000+ tok/s generation with 128k context (text only). Community notes that human code review is now the bottleneck.

MiniMax M2.5: 230B-Parameter MoE, Cheap Long-Run Agent Model

230B total parameters with 10B active. SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3%, BrowseComp 76.3. Near-Opus coding performance at roughly 1/13th the price (~$1/hour at 100 tok/s; four always-on instances ≈ $10K/year). Available on OpenRouter, NetMind, and others.

Zhipu GLM-5: Top Open-Weight Model

Scored 50 on the Intelligence Index (claimed #1 open weight, lowest hallucination rate) and #1 among open models on Arena text and code leaderboards. Rumored ~744B total / ~40B active parameters, 28.5T training tokens, DeepSeek-style sparse attention, GGUF available for local inference.

DeepSeek V4 Teased

Officially previewed for release around February 17 with 1M-token context, positioned as a cost-effective alternative to closed models like Opus and Codex.

Tiny QED-Nano: 4B Proof Model

Natural-language theorem-proving model approaching larger models on IMO-ProofBench via heavy test-time compute (1M+ token agent scaffolds) and rule-to-reward RL post-training; weights and training details planned for open sourcing.

Agents & Tooling

Agent coding race: MiniMax M2.5 and GLM-5 both target agentic workflows — task decomposition, long executions, multi-file refactors — with adoption by Cline, YouWare, Verdent, and other IDEs/agent frameworks.
Cursor long-running agents: Cursor shipped long-running agents for large refactors; paired with Codex Spark, generation is near-instant, though pricing/quota changes are drawing complaints.
A2A protocol: Andrew Ng promoted DeepLearning.AI's Agent2Agent protocol for inter-agent discovery and invocation, interoperating with IBM ACP, Google ADK, LangGraph, and MCP.
Mooncake: Moonshot/Tsinghua's KV-cache serving backend joined the PyTorch ecosystem, addressing the LLM inference "memory wall" with prefill/decoding disaggregation, cross-instance cache reuse, and elastic expert parallelism; works as a backend for SGLang, vLLM, TensorRT-LLM.
Lightweight tools trend: Communities increasingly pair specialized components (e.g., no-API-key Google Search MCP built on Chromium, SigLIP2 for bulk image tagging) instead of multimodal LLMs for everything.
Traces platform: New platform for sharing and browsing full coding-agent session traces exported from Claude Code, Codex, Gemini, and Cursor.
Tool updates: Aider v0.86.2 released; Windsurf added Opus 4.6, SWE-1.5, and GPT-5.3-Spark to its Arena; OpenRouter onboarded MiniMax M2.5 and GLM-5.

Infrastructure & Hardware

torchao v0.16.0: Adds MXFP8 MoE training components for expert parallelism, removes legacy quantization options for a leaner, ABI-stabler stack.
GPU MODE kernel experiment: Plans to spend $20–30K on 4–5 days of compute (late February) testing CUDA/Triton kernel generation with Qwen3/GLM4.7 Flash models, using Kernelbot and Flashinfer-bench, building SFT baselines before RL.
TraceML: Open-source real-time observability for PyTorch DDP — per-rank step times and jitter with a one-line hook, exposing slow stragglers.
RTX 3060 12GB stock-up: ~$200 clearance cards make cheap 24GB (two-card) local LLM/CUDA setups; community sees these as budget AIGC workhorses.
NCU/ROCm details: Community explainers on NCU's Local(57) field (local memory instructions, often register spilling), tcgen05 usage, and AMD Quick Reduce working on MI250X/CDNA2 despite official guards limiting it to MI300.

Research & Methods

Anthropic "Introspection": A monitoring network watches the main model's activations/weights for anomalies. Critics (e.g., Unsloth) call it redundant distribution-shift detection; supporters note practical safety value in detecting subtle manipulation signs.
BlendFER-Lite: LSTM + MediaPipe 3D expression parameters for facial expression recognition on edge devices; 71% accuracy on FER2013 with low compute; code/models on Hugging Face.
Rank-1 LoRA vs full RL: Thinking Machines experiments show rank-1 LoRA fine-tuning can match full-model RL fine-tuning on some reasoning tasks; follow-up work debates whether it's just in-context learning.
GoodfireAI RLFR: Reinforcement Learning from Feature Rewards — extract interpretable features first, then build reward functions from them for open-ended RL tasks.
Emergence debate: Nick Bostrom's new paper (A Theory of Emergent Behaviour) examines advanced AI risk philosophically, while the "are emergent abilities a mirage?" debate (arXiv:2304.15004) continues.

Products & Applications

Simile: Raised $100M (investors include Karpathy and Fei-Fei Li) to build "societal digital twins" using generative agents for pre-launch decision testing (earnings call rehearsals, policy changes).
ChatGPT medical catch: Multiple Reddit users report ChatGPT urged immediate care that surfaced pulmonary embolism, heart attack, and shingles — useful as triage/second opinion, not diagnosis.
Chinese-model coding feedback: Kimi K-2.5 praised for multimodal browsing and web tasks but criticized on complex codebases; GLM-5 and DeepSeek V3.2 steadier for tool calling and long context. Users commonly mix models.
Weave Robotics Isaac 0: $8,000 (or $450/month) laundry-folding home robot, first deliveries February 2026, San Francisco Bay Area only.
Everyday AI use cases: From creative writing to automated resume blasts — sticky personal use cases remain highly practical.

Business & Industry

Anthropic raises at ~$380B valuation: Reported annualized revenue around $14B; Claude Code ARR reportedly doubled to $2.5B this year.
Perplexity quota backlash: Pro users' Deep Research quietly cut to 20–50 runs/month, triggering cancellations and migrations to Google AI Pro or self-built stacks.
xAI power questions: Unverified reports claim partially unpermitted gas turbines plus grid power for Grok training/inference — highlighting how power-constrained frontier AI competition has become.
Anthropic covers electricity price hikes: The company will absorb electricity cost increases in some regions to keep enterprise pricing stable — evidence that power costs now directly hit AI margins.
Local vs cloud split: Communities diverge between budget local stacks (3060s, LM Studio, Unsloth) for privacy/control and cloud tools (Cursor, Claude Code, Copilot) for convenience.

Policy, Governance & Safety

Jailbreak escalation: BASI community reports patched Claude Code jailbreaks, new DAN prompts for GPT-5.2/Gemini 3 Fast, and automated red-team systems ("HAIL MARY" built on Manus) probing safety policies 24/7.
"AI psychosis" cases: Psychiatry podcast covers cases of delusion amplification in vulnerable users during long, emotionally intense chatbot sessions — arguing for usage-guardrail-style designs.
RLHF and sycophancy debate: Some argue RLHF rewards persuasiveness over truthfulness, training better confabulators; others blame training-objective design rather than inherent tendencies.
KOKKI v15.5 pattern: A proposed dual-output mode (Draft + Audit) where models self-audit, trading token cost and latency for observability; the community argues true guarantees require deterministic systems, so transformers can only aim for auditable, bounded error distributions.

---

📌 Source: Easy AI Daily