Easy AI Daily | February 11, 2026

A structured digest of AI industry news for February 11, 2026, translated and summarized from the Easy AI Daily (zhichai.net).

Key points

Models & Capabilities

• Qwen-Image-2.0 (Alibaba): A 7B unified text-to-image + image editing model, natively 2K resolution, ~1K token prompts, strong typography, Chinese calligraphy, infographics, and multi-panel comic consistency. Shrunk from 20B to 7B vs. the previous generation—faster inference and easier local deployment; community awaits open weights. Official blog | Reddit discussion
• Seedance 2.0 (ByteDance): Text-to-video model going viral for natural motion and improved body/cloth physics; high-quality anime fight scenes and realistic 1v1 basketball clips; currently ~15 seconds per clip. Petapixel coverage
• Kimi K2.5 + Agent Swarm (Moonshot): Up to ~100 sub-agents per task and 1,500 tool calls; claimed ~4.5x speedup via parallel execution. Community demos include a storyboard-to-video pipeline feeding Seedance 2. Announcement
• Claude Opus 4.6 (Anthropic): #1 on Arena text and code leaderboards; praised for one-shotting complex UI and long-form writing, but slower and token-hungry. Safety researchers criticized reliance on internal staff surveys for high-risk capability thresholds. RSP critique
• OpenAI Deep Research → GPT-5.2: Adds external data connectors and progress controls, positioning it as a long-running research agent. OpenAI announcement
• Gemini 3 Pro: New checkpoint spotted in A/B testing; paid users report regressions (hallucinations, weaker coding) while others prefer its personality and citations. TestingCatalog
• Underrated open multimodal models: GLM-OCR (documents), MiniCPM-o-4.5 (on-phone GPT-4o-like), InternS1 (scientific figures/papers) — all free for commercial use.
• GLM-4.7-Flash-GGUF: Now a top-downloaded model on Unsloth; local LLM momentum in the Chinese community.
• IsoDDE (DeepMind/Isomorphic Labs): Claims roughly doubled performance on protein/molecular structure benchmarks, with antibody binding-site predictions beyond physical simulation; few architecture details released. Isomorphic Labs

Agents & Tooling

• OpenAI Responses API: Server-side context compaction, hosted networked containers for code execution, and Skills as first-class citizens.
• LangChain deepagents v0.4: Pluggable sandbox backends (Modal/Daytona/Runloop), OpenAI Responses as default backend; frames sandboxes as tools while agents stay resumable.
• Coding agent UX: VS Code/Copilot adding worktrees, MCP apps, slash commands; multi-model workflows (Opus 4.6 + GPT-5.3-Codex + Gemini 3 Pro) with cross-review sub-agents.
• Claude Code hidden flag --sdk-url: Turns the CLI into a WebSocket client for browser/mobile frontends, enabling remote-IDE setups.
• Electric SQL "Configurancy": Configuration-driven, replayable workflows with budgets and guardrails to make agent-written code maintainable. Blog
• LMArena: New PDF-prompt comparisons; academic partnerships program funding up to $50,000 per project.

Infrastructure & Hardware

• Unsloth: Up to 12x MoE training speedup, ~35% memory savings via custom Triton kernels and torch._grouped_mm; small MoE trainable on <15GB VRAM. Docs
• vLLM in production: AI21 doubled burst throughput via config + queue-based autoscaling; postmortem on a rare ~1-in-1000 garbled-output bug traced to request-classification timing under memory pressure.
• Distributed inference: Kimi K-2.5 (658GB) run across 4 Thunderbolt-linked Mac Studios via MLX Distributed; Qwen3Next Q4 at ~40 tok/s on an AMD H395 AI MAX laptop.
• Nubank: Hiring CUDA engineers to train in-house models on B200s. Paper
• Modular acquires BentoML: Plans a write-once, run-anywhere inference stack across NVIDIA, AMD, and future accelerators; BentoML stays Apache 2.0.

Research & Methods

• iGRPO: Two-stage RL — sample drafts, select the best via unified scoring, then train to beat it; no critic needed; outperforms GRPO on 7B/8B/14B families.
• Self-verification & ConceptLM: Self-check-and-repair yields better reasoning with fewer tokens; ConceptLM quantizes hidden states into a concept vocabulary for next-concept prediction.
• Ganguli scaling theory: Conditional entropy decay and token-correlation decay in natural language can predict scaling-law exponents in data-constrained regimes.
• Architecture leaks from source: GLM-5 reportedly ~740B total / ~50B active with MLA-style attention and 200K context; Qwen3.5 rumored to alternate Gated DeltaNet linear attention with standard attention in an MoE hybrid SSM-Transformer design.
• Generative Meta-Model (Radford et al.): Diffusion trained on 1B LLM activations for on-manifold steering; promising for controllability and interpretability. Project page
• Model self-reflection study: Llama3.1/Qwen2.5-32B spontaneously coin words (e.g., "loop," "mirror") describing their activation patterns, correlating with real autocorrelation/spectral features. DOI

Products & Applications

• Perplexity Pro: Quiet quota cuts (Deep Research to 20/month, uploads to 50/week) triggered subscriber backlash and bot-heavy support complaints.
• Cursor: Composer 1.5 at 50% off, but users report instability, forced Auto-model routing, slow queues, and opaque billing.
• Arena image/video revamp: Clustered 4M prompts into scenario-specific leaderboards (~15% noise removed); Veo 3.1 tops Video Arena.
• P402.io: An OpenRouter middleware for automatic model selection and cost accounting with USDC/USDT payments (1% fee).
• AuditAI (open source): Corrective RAG (LangGraph) for NIST CSF 2.0 compliance auditing with mandatory evidence citations; evaluated with Llama 3.3 70B + Groq. GitHub

Industry & Companies

• Qwen3-Coder-Next gaining local-LLM fans as a strong general-purpose small model; community debates whether $5k–10k local machines beat cloud subscriptions long-term.
• Salesforce executive exodus: Slack/Tableau CEOs, president, and CMO departing to OpenAI/AMD, fueling AI-strategy speculation.
• Cloudflare surpassed $2B annual revenue; Vercel CEO offered to cover a developer's $46K HTML-rendering bill after a viral cost incident.
• a16z invests in Shizuku AI Labs (Japan): AI companion characters blending Japanese character design with frontier models.

Policy, Governance & Safety

• OpenClaw architecture criticized for enabling indirect authorization jailbreaks via loose permissions; embedding whitelists + syntactic constraints suggested as mitigations. Repo
• GPT-5.2 / Opus 4.6 jailbreak arms race: ENI and Glossopetrae-style techniques evolving; some teams monetizing red-team services.
• Discord ID verification for some channels draws developer backlash amid IPO-compliance speculation.
• OpenAI tests ads in ChatGPT in search/info scenarios, raising concerns about ranking bias and user trust. OpenAI
• KOKKI Agent-Auditor Loop: Draft-then-audit architecture (cross-model auditing recommended) reportedly reduces hallucination rates in early tests.

📌 Source: Easy AI Daily (zhichai.net)