Easy AI Daily Digest - February 11, 2026: Qwen-Image-2.0, Seedance 2.0, Kimi K2.5, Claude Opus 4.6 and More

Models & Capabilities

Alibaba releases Qwen-Image-2.0: 7B unified text-to-image + editing Qwen's Qwen-Image-2.0 is a 7B-parameter unified model supporting text-to-image generation and image editing, with native 2K resolution and prompts up to ~1K tokens. It handles high-quality typography, Chinese calligraphy, infographics, and character consistency across comic panels. Shrunk from 20B to 7B versus the previous generation, it runs faster and is better suited to local deployment; the community awaits open weights.

ByteDance Seedance 2.0 feels like a generational leap Seedance 2.0 flooded Reddit and X, dubbed the end of the "Will Smith spaghetti" meme. Community testing shows natural motion and clearly improved body/cloth physics; it generates high-quality anime fight scenes and 1v1 basketball videos. Current clips run ~15 seconds, with users clamoring for longer durations.

Moonshot Kimi K2.5 and Agent Swarm: massive parallel sub-agents Moonshot released Kimi K2.5 and Agent Swarm, which can schedule up to ~100 sub-agents and 1,500 tool calls per task, claiming ~4.5x speedup from parallel over serial execution. One user built a full storyboard-to-video pipeline feeding huge Excel storyboards from K2.5 into Seedance 2.

Anthropic Claude Opus 4.6: #1 on Arena, big safety-gate controversy Claude-opus-4.6-thinking ranked #1 on both Arena text and coding leaderboards. Practitioners report it "one-shots" complex UI designs and long-form writing far better than 4.5, though it is slower and more token-hungry. Safety researchers criticize Anthropic for relying mainly on internal staff surveys for high-risk capability gates, lacking hard-metric evaluations.

OpenAI Deep Research switches to GPT-5.2, adds connectors and progress control OpenAI upgraded Deep Research's underlying model to GPT-5.2 and added external data connectors and progress controls, leaning into a "long-duration research agent" design. Some wonder why not 5.3; others speculate a Codex variant shipped first.

Gemini 3 Pro new checkpoint spotted in testing; paid tier gets mixed reviews A new Gemini 3 Pro checkpoint appeared in A/B testing. Meanwhile many paid users on Reddit complain about regression: hallucinations, irrelevant padding, and coding ability beaten by Copilot and Cursor; other users prefer its "engineer + librarian" personality and heavier citation use.

Underrated open multimodal models: GLM-OCR, MiniCPM-o-4.5, InternS1 A community roundup highlights GLM-OCR for document recognition, MiniCPM-o-4.5 running a "GPT-4o-like" model on phones, and InternS1 excelling at scientific charts and paper understanding—all free for commercial use.

Zhipu GLM-4.7-Flash-GGUF tops Unsloth downloads The GGUF version of GLM-4.7-Flash became one of the most-downloaded models on Unsloth, signaling rising local-LLM momentum in the Chinese community.

DeepMind/Isomorphic Labs announce IsoDDE, claiming to far surpass AlphaFold 3 Isomorphic Labs' IsoDDE technical report claims doubled performance on protein/molecule structure benchmarks, with antibody binding-site and affinity predictions beating physical simulations. If it holds up, in-silico drug discovery gets another speed boost, though architecture details remain scarce.

Agents & Tooling

• OpenAI Responses API embraces long-running agents: server-side automatic context compaction, official hosted networked containers for code execution, and Skills promoted to first-class citizens (e.g., built-in spreadsheet skills).
• LangChain deepagents v0.4: pluggable sandbox backends (Modal/Daytona/Runloop), improved summarization/compression, and OpenAI Responses as default—favoring "sandbox as a tool, agent stays resumable."
• Coding agent UX evolves fast: VS Code/Copilot adds worktrees, MCP apps, slash commands; practitioners run Claude Opus 4.6, GPT-5.3-Codex, and Gemini 3 Pro in parallel with sub-agents cross-reviewing each other. GPT-5.3-Codex's official VS Code rollout was briefly paused but is seen as more token-efficient for large workflows.
• Claude Code hidden flag --sdk-url discovered: turns the local CLI into a WebSocket client for browser/mobile front ends—effectively an official remote-IDE protocol, enabling fully automated tmux + multi-sub-agent setups.
• Electric SQL proposes "Configurancy": treat agent-written code as a configuration-driven, rerunnable pipeline with budgets and guardrails—more like CI/CD than one-off scripts.
• Arena adds PDF prompts for model comparison and launches an academic partnerships program funding projects up to $50,000.

Infrastructure & Hardware

• Unsloth 12x MoE training speedup: custom Triton kernels and torch._grouped_mm cut VRAM ~35% and enable small MoE training on <15GB consumer GPUs; also shipped long-context RL tutorials.
• vLLM production war stories: AI21 doubled burst-traffic throughput via config + queue-based autoscaling, and debugged a rare bug corrupting ~0.1% of responses, traced to request-classification timing under memory pressure.
• Distributed inference on Macs: Kimi K-2.5 (658GB) run across 4 Thunderbolt-linked Mac Studios via MLX Distributed, claiming near-linear throughput scaling. Also: Qwen3Next Q4 at ~40 tok/s on an AMD H395 "AI MAX" laptop with 96GB unified memory.
• Nubank hiring CUDA experts to train its own foundation models on B200s (team includes the Liger Kernel author; see arXiv:2507.23267).
• Modular acquires BentoML: combining BentoML deployment with MAX/Mojo for "write once, run anywhere" inference across NVIDIA, AMD, and future accelerators; BentoML remains Apache 2.0.

Research & Methods

• iGRPO: two-stage GRPO variant—sample drafts, pick the best via unified scoring, then train the model to beat that draft—no critic or textual feedback needed; reportedly outperforms GRPO across 7B/8B/14B families.
• Self-verification & ConceptLM: models answer, then self-check and correct, achieving better reasoning with fewer tokens; ConceptLM quantizes hidden states into a "concept vocabulary" for next-concept prediction, gaining further via continued pretraining.
• Ganguli: language statistics alone—conditional entropy decay and token correlation decay—can predict scaling-law exponents in data-constrained regimes.
• Source-digging reveals architectures: GLM-5 reportedly ~740B total / ~50B active params with DeepSeek-V3-style MLA and sparse indexing for 200K context; Qwen3.5 rumored as a hybrid SSM-Transformer alternating Gated DeltaNet linear attention with standard attention plus MoE with shared experts.
• Generative Meta-Model (Radford et al.): diffusion trained on 1B LLM activations enabling on-manifold steering in activation space—promising for controllable generation and interpretability (project page).
• Model self-reflection: Llama3.1 and Qwen2.5-32B spontaneously invent words (e.g., "loop", "mirror") describing their own activation patterns, with frequencies correlating with real autocorrelation/spectral features (paper).

Products & Applications

• Perplexity Pro quietly cuts quotas: Deep Research limited to 20/month, file uploads to 50/week—paid users mass-cancel and criticize bot-heavy customer support.
• Cursor: Composer 1.5 at 50% off, but users report forced Auto-model switches, disconnects, slow queues, and opaque Auto billing.
• Arena image/video leaderboards overhauled: 4M user prompts clustered into scenario-specific leaderboards (commercial design, 3D modeling, etc.), ~15% noise filtered; Video Arena moved from Discord to web. Veo 3.1 tops Video Arena.
• P402.io: an OpenRouter middleware layer tracking per-model request costs, comparing price/performance (e.g., Opus 4.6 vs Sonnet 4.5), suggesting replacements; supports USDC/USDT with 1% fees.
• AuditAI: open-source corrective-RAG (LangGraph-based) tool auditing security policies against NIST CSF 2.0, with mandatory evidence citations, evaluated with Llama 3.3 70B + Groq.

Industry & Company News

• Qwen keeps winning local-LLM fans: Qwen3-Coder-Next praised as "smartest small model"—strong as a general assistant despite the Coder name.
• Local LLM economics debated: $5k–10k hardware vs cloud subscriptions—capability still lags cloud, but fast small-model progress appeals to privacy-sensitive teams with stable workloads.
• Salesforce exodus: Slack/Tableau CEOs, president, and CMO depart for OpenAI and AMD, read as a referendum on Salesforce's AI strategy.
• Cloudflare tops $2B revenue; a developer's $46K Vercel bill for Jmail HTML rendering drew the Vercel CEO's personal offer to help and cover costs.
• a16z leads investment in Japan's Shizuku AI Labs, betting on "2D anime AI companion" products combining Japanese character design with frontier models.

Policy, Governance & Safety

• OpenClaw architecture enables indirect privilege-escape jailbreaks: loose permissions and weak system prompts let agents bypass intended boundaries; suggested mitigations include embedding whitelists plus syntactic constraints.
• GPT-5.2 vs Opus 4.6 jailbreak arms race: ENI attacks partially succeed on 4.6; "Glossopetrae"-style virtual-world-language bypasses keep evolving; some teams now sell these as paid red-team services.
• Discord demands ID uploads from some users, drawing developer backlash—speculation ranges from pre-IPO compliance to mass data collection.
• OpenAI tests ads in ChatGPT for some search/info queries, sparking debate over commercial influence on model outputs and user trust.
• KOKKI Agent-Auditor Loop (v15.5): a proposed draft-then-audit pipeline (Audit(Draft(input))), with cross-model auditing (e.g., GPT drafts, Claude audits) reportedly more robust than single-model self-review.

📌 Source: Easy AI Daily