Easy AI Daily News | February 13, 2026: Gemini 3 Deep Think V2, GPT-5.3-Codex-Spark, MiniMax M2.5, GLM-5

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for February 13, 2026 rounds up the biggest AI industry developments. Google released Gemini 3 Deep Think V2 scoring 84.6% on ARC-AGI-2 (ARC-certified) and launched Aletheia, a math-specialized agent scoring 100% on IMO 2024 tasks. OpenAI partnered with Cerebras to preview GPT-5.3-Codex-Spark, a coding model generating 1000+ tokens/second with 128k context. MiniMax shipped M2.5, a 230B-parameter (10B active) MoE agent model scoring 80.2% on SWE-Bench Verified at roughly 1/13 the cost of Claude Opus. Zhipu's open-weight GLM-5 topped open-model leaderboards with an Intelligence Index score of 50, while DeepSeek teased V4 with 1M-token context for around February 17. On infrastructure, Mooncake (Moonshot/Tsinghua KV-cache backend) joined the PyTorch ecosystem, and torchao v0.16.0 added MXFP8 MoE support. Business news includes Anthropic's ~$380B valuation round and Perplexity cutting Deep Research quotas for Pro users.

Easy AI Daily | 2026-02-13

A daily digest of AI industry news curated by zhichai.net, covering models, agent tooling, infrastructure, research, products, business, and safety.

Key points

Google Gemini 3 Deep Think V2: Reasoning mode for paid Gemini users (early access via Vertex AI/Gemini API). ARC-AGI-2: 84.6% (ARC-Prize certified), HLE 48.4%, Codeforces Elo 3455, olympiad-level physics/chemistry and IMO-level math. Used for semiconductor materials design, math paper review, and sketch-to-CAD/STL workflows; per-task cost on ARC reportedly cut to ~18% of the previous version.
Google Aletheia: DeepMind's math-research agent scores 91.9% on IMO-ProofBench Advanced and 100% on IMO 2024 tasks, using a generator-plus-verifier architecture; methods to be extended to physics and CS.
OpenAI GPT-5.3-Codex-Spark: Research preview for ChatGPT Pro Codex (app/CLI/IDE), built with Cerebras. Claims 1000+ tok/s generation, 128k context, text-only. Community notes that human code review is now the bottleneck.
MiniMax M2.5: 230B total / 10B active MoE for long-horizon agents. SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3%, BrowseComp 76.3; near-Opus performance at ~1/13 the price (~$1/hour at 100 tok/s). Available on OpenRouter, NetMind, and more.
Zhipu GLM-5: Strongest open-weight model on several leaderboards — Intelligence Index 50 (open-weight #1, lowest hallucination rate), #1 open model on Arena text and code. Rumored ~744B total / ~40B active params, 28.5T training tokens, DeepSeek-style sparse attention; GGUF quants available.
DeepSeek V4 teaser: Expected around Feb 17 with 1M-token context, positioned as a cost-effective alternative to closed models like Opus and Codex.
Tiny QED-Nano: A 4B natural-language theorem-proving model approaching large-model benchmarks via heavy test-time compute (1M+ token agent scaffolds) and rule-based RL post-training; weights planned to be open-sourced.

Agents and tooling

MiniMax M2.5 and GLM-5 mark a new round of "agentic coding model" competition; GLM-5 is being integrated into Cline, YouWare, Verdent, and other IDEs/agent frameworks.
Cursor launched a long-running Agent; paired with GPT-5.3-Codex-Spark, code generation feels near-instant, but pricing/quota changes (Composer 1.5 price hikes) are drawing complaints.
Andrew Ng promoted DeepLearning.AI's Agent2Agent (A2A) protocol for inter-agent interoperability, connecting IBM ACP, Google ADK, LangGraph, and MCP.
Mooncake (Moonshot AI × Tsinghua) joined the PyTorch ecosystem: a KV-cache serving backend addressing the LLM inference memory wall, with prefill/decoding disagg, cross-instance cache reuse, and elastic expert parallelism; works as a backend for SGLang, vLLM, TensorRT-LLM.
Community trend toward specialized tools over monolithic LLMs: no-API-key Google Search MCP (Chromium-based, with YouTube transcripts and local OCR), SigLIP2 for large-scale image tagging.
Traces: a new platform for sharing/browsing full coding-agent session traces exported from Claude Code, Codex, Gemini, Cursor.
Aider v0.86.2 shipped; 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 and hardware

torchao v0.16.0: adds MXFP8 MoE training components for expert parallelism; removes legacy configs for a leaner, ABI-stable quantization stack.
GPU MODE plans to spend $20–30k of compute over 4–5 days in late February testing AI-generated CUDA/Triton kernels with Qwen3/GLM4.7 Flash, using Kernelbot, FlashInfer-bench, NCU, and Compute-Sanitizer.
TraceML: open-source one-line hook for real-time PyTorch DDP observation, exposing per-rank step timing and jitter.
Community is bulk-buying Zotac RTX 3060 12GB cards (~$200) for budget local LLM rigs — 24GB VRAM for two cards.
NCU field explanations (e.g., Local(57) local-memory instructions from register spills), tcgen05 usage notes, and AMD Quick Reduce enablement questions on MI250X/CDNA2.

Research and methods

Anthropic "Introspection": an auxiliary network monitoring the main model's activations/weights; critics (e.g., Unsloth) call it distribution-shift detection, supporters see safety value.
BlendFER-Lite: LSTM + MediaPipe facial expression recognition hitting 71% on FER2013 with small compute, suited to edge devices; code on Hugging Face.
Rank-1 LoRA: Thinking Machines experiments show rank-1 LoRA fine-tuning can match full RL fine-tuning on some reasoning tasks.
GoodfireAI RLFR: Reinforcement Learning from Feature Rewards — interpretable features as more stable reward functions for open-ended tasks.
Ongoing debate on emergent abilities: Nick Bostrom's new paper (arXiv:2511.10643) vs. the "mirage" argument (arXiv:2304.15004).

Products and applications

Simile raised $100M (investors include Karpathy and Fei-Fei Li) for generative-agent-based "societal digital twins" to test business/policy decisions before rollout.
Multiple Reddit users report ChatGPT urging immediate medical care that led to diagnoses like pulmonary embolism and heart attack — useful as a second-opinion/triage layer, not a doctor replacement.
Chinese-model coding feedback: Kimi K-2.5 praised for web browsing and site cloning; GLM-5 and DeepSeek V3.2 more stable in tool calling and long context; users commonly mix models.
Weave Robotics Isaac 0: an $8,000 (or $450/month) laundry-folding home robot, first deliveries February 2026, San Francisco Bay Area only.

Industry and business

Anthropic confirmed a new round at ~$380B valuation; reported annualized revenue ~$14B; Claude Code ARR said to have doubled to $2.5B. Anthropic also announced it will cover electricity price increases for enterprise customers.
Perplexity silently cut Pro Deep Research quotas to 20–50/month, triggering cancellations.
Unverified reports claim xAI's compute partially relies on unpermitted gas turbines plus grid power.
Community debate continues on local vs. cloud LLM setups: hobbyists favor 3060s + LM Studio/Unsloth; enterprises lean on Cursor, Claude Code, Copilot.

Policy, governance, and safety

BASI jailbreaking community reports patched Claude Code exploits, new DAN scripts for GPT-5.2/Gemini 3 Fast, and an automated red-team system ("HAIL MARY") built with Manus.
A psychiatry podcast covers emerging "AI psychosis" cases of delusion amplification via LLM chatbots.
Debate on whether RLHF rewards persuasiveness over truthfulness; OpenAI community proposes KOKKI v15.5 dual-output (Draft + Audit) for auditable model responses.

---

📌 Source: Easy AI Daily (zhichai.net)