Easy AI Daily Digest — February 11, 2026

A curated translation of the day's AI industry news, originally published in Chinese on zhichai.net.

Models & Capabilities

Alibaba releases Qwen-Image-2.0: 7B unified text-to-image + editing

Alibaba's Qwen team launched Qwen-Image-2.0, a 7B-parameter model unifying 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 in multi-panel comics. Shrunk from 20B (previous generation) to 7B, it's faster for inference and easier to deploy locally; the community awaits open weights.

• Official blog
• Reddit discussion

ByteDance Seedance 2.0: text-to-video feels "a generation upgraded"

Seedance 2.0 flooded Reddit and X, dubbed the end of the "Will Smith spaghetti" meme era. Community tests show natural motion, markedly better body and cloth physics, and high-quality anime fight scenes and 1v1 basketball videos. Clips run ~15 seconds; users are asking for longer durations.
• Petapixel coverage
• Will Smith spaghetti example
• Anime fight compilation
• 1v1 basketball discussion

Moonshot Kimi K2.5 and "Agent Swarm": massively parallel sub-agents

Moonshot released Kimi K2.5 and Agent Swarm, able to dispatch up to ~100 sub-agents and 1,500 tool calls per task, claiming ~4.5x speedup via parallel execution. One community member generated a huge Excel storyboard with K2.5, then fed it to Seedance 2 for a full storyboard-to-video pipeline.
• Agent Swarm announcement
• K2.5 + Seedance 2 storyboard case
• Baseten benchmarks (TTFT/TPS)

Anthropic Claude Opus 4.6: #1 on leaderboards, but safety-gate controversy

Claude-opus-4.6-thinking took #1 on both Arena text and code leaderboards. Practitioners report it one-shots complex UI designs and long-form writing noticeably better than 4.5, though it's slower and more token-hungry. Safety researchers criticized that Anthropic appears to judge high-risk capability thresholds mainly via internal employee surveys, lacking hard evaluation metrics.
• Arena leaderboard
• Opus 4.5 vs 4.6 UI comparison
• RSP safety-gate criticism

OpenAI Deep Research switches to GPT-5.2, adds connectors and progress control

OpenAI upgraded ChatGPT's Deep Research underlying model to GPT-5.2, adding external data connectors and progress-bar controls—positioning it as a long-duration research agent. Some wonder why not 5.3; others guess a Codex variant shipped first with the main model to follow.
• OpenAI announcement

New Gemini 3 Pro checkpoint spotted in testing; paid-tier reception mixed

A new Gemini 3 Pro checkpoint was found in A/B testing, likely a refinement of Gemini 3. Meanwhile many paid users on Reddit complain about regression: hallucinations, irrelevant padding, and coding ability outclassed by Copilot and Cursor; others prefer its personality and richer citations.
• A/B test leak
• Regression complaints

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 GPT-4o-like capabilities on phones, and InternS1 excelling at scientific charts and paper understanding—all claimed free for commercial use.
• Model roundup

Zhipu GLM-4.7-Flash-GGUF tops Unsloth downloads

Zhipu noted GLM-4.7-Flash's GGUF became one of the most-downloaded models on Unsloth, signaling rising local-LLM momentum, boosted by Unsloth's MoE training optimizations.
• Z.ai announcement

DeepMind/Isomorphic Labs unveil IsoDDE: claimed to far surpass AlphaFold 3

Isomorphic Labs' IsoDDE technical report claims roughly doubled performance on protein and molecular structure benchmarks, with antibody binding-site and affinity predictions reportedly more accurate than physical simulation. If it holds up, in-silico drug discovery could accelerate another notch—though few architecture details are public.
• Technical report intro
• Demis Hassabis comment

Agents & Tooling

OpenAI Responses API embraces long-running agents

Three new capabilities: server-side automatic context compaction, official hosted networked containers (run code), and Skills as first-class citizens (built-in spreadsheet skill, etc.). Combined with the Deep Research upgrade, this signals research/computer-use agents becoming formal product surfaces.
• OpenAI dev update

LangChain deepagents v0.4: pluggable sandboxes, OpenAI Responses default

LangChain added pluggable sandbox backends (Modal/Daytona/Runloop) in deepagents v0.4, enhanced summarization/compaction, and made OpenAI Responses the default backend—siding with the "sandbox as tool, agent stays resumable" architecture.
• Harrison Chase architecture analysis
• deepagents v0.4 update

Coding agent UX evolves fast: multi-model collaboration becomes standard

VS Code/Copilot keeps adding agent features (worktrees, MCP apps, slash commands). Practitioners run Claude Opus 4.6, GPT-5.3-Codex, and Gemini 3 Pro in parallel with sub-agents reviewing each other. GPT-5.3-Codex's official VS Code launch briefly stalled but is seen as more token-efficient for large workflows.
• Multi-model code review example

Hidden Claude Code flag discovered: turn CLI into a browser/WebSocket client

The --sdk-url flag in Claude Code CLI lets it act as a WebSocket client for custom backends—an unofficial "remote IDE protocol." People are building fully automated tmux + multi-sub-agent coding setups.
• Stan Girard demo

Electric SQL's "Configurancy": configuration-driven agent coding

Electric SQL argues maintainable agent-written code requires designing configuration and structural constraints up front, not freewheeling prompts. Their Configurancy/OpenProse concepts express workflows as rerunnable, budgeted, guardrailed pipelines—closer to CI/CD than one-off scripts.
• Configurancy post

Arena adds PDF prompts and an academic evaluation fund

LMArena now supports prompts with attached PDFs to compare long-document extraction, summarization, and reasoning. Its new academic partnerships program funds projects up to $50,000 each.
• Image Arena improvements
• Academic Partnerships Program

Infrastructure & Hardware

Unsloth: 12x faster MoE training, 35% less VRAM, long-context RL

Using custom Triton kernels and torch._grouped_mm, Unsloth speeds MoE training up to 12x with ~35% VRAM reduction—small MoEs trainable on <15GB consumer GPUs. They also released long-context RL tutorials and local LLM + Claude Code/Codex integration guides.
• Faster MoE docs
• Long-context RL guide

vLLM production lessons: throughput tuning and a rare 0.1% garbled-output bug

AI21 doubled burst-traffic throughput via configuration plus queue-based autoscaling; another post traces a bug where ~1 in 1,000 requests returned garbled text to a request-classification timing issue under memory pressure.
• Throughput tuning
• Garbled-output postmortem

Distributed inference tricks: Kimi K-2.5 on 4 stacked Mac Studios

Using MLX Distributed, one user split Kimi K-2.5's 658GB footprint across 4 Thunderbolt-linked Mac Studios, claiming near-linear throughput scaling (at linearly scaling cost). Another ran Qwen3Next Q4 at ~40 tok/s on an AMD H395 "AI MAX" laptop with 96GB unified memory.
• Mac cluster share

Nubank hires CUDA experts to train in-house foundation models on B200s

The Brazilian fintech is recruiting CUDA/kernel engineers via the GPU MODE community, with Liger Kernel authors on the team and a recent paper published.
• Hiring post and paper

Modular acquires BentoML: "write once, run anywhere" inference stack

Modular plans to combine BentoML deployment with MAX/Mojo so the same code runs on NVIDIA, AMD, and future accelerators. BentoML remains Apache 2.0 open source.
• Acquisition announcement
• AMA thread

Research & Methods

iGRPO: RL improvement without a critic, competing against your own best draft

iGRPO extends GRPO in two stages: sample multiple drafts, pick the best via unified scoring, then train the model to beat that draft—no critic or textual feedback needed. Reported to outperform GRPO across 7B/8B/14B model families.
• Method thread

Self-verification & ConceptLM: thinking more clearly with fewer tokens

"Learning to Self-Verify" lets models answer, then check and correct themselves, achieving better reasoning with fewer tokens. ConceptLM quantizes hidden states into a "concept vocabulary" for next-concept prediction, claiming continued pretraining gains.
• Self-verification discussion

Ganguli: predicting data-constrained scaling law exponents from language statistics alone

Surya Ganguli shared theoretical results showing that conditional entropy decay over context length and token-correlation decay can predict scaling law exponents in data-limited settings—a theoretical yardstick for how much data is actually needed.
• Thread

Architecture archaeology: GLM-5 and Qwen3.5 structures leaked

Developers digging through repos and logs claim GLM-5 has ~740B total / ~50B active parameters with MLA attention and sparse indexing for 200K context; Qwen3.5 is speculated to be a hybrid SSM-Transformer alternating Gated DeltaNet linear attention with full attention, plus MoE expert sharing.
• GLM-5 leak
• Qwen3.5 leak

"Generative Meta-Model": diffusion directly in activation space

A short paper trains a diffusion model on 1B LLM activations for on-manifold steering—using a small model to manipulate a large model's internal states. The Eleuther community sees promise for controllable generation and interpretability.
• Project page

Model self-reflection study: Llama3.1/Qwen2.5 invent words for their internal states

In self-dialogue experiments on Llama3.1 and Qwen2.5-32B, models spontaneously coined vocabulary (e.g., "loop," "mirror") describing their activation patterns, with frequencies correlating strongly with real autocorrelation/spectral features.
• Paper

Products & Applications

Perplexity Pro quietly cuts quotas; users cry "bait and switch"

Without notice, Perplexity limited Pro's Deep Research to 20 runs/month and file uploads to 50/week, triggering unsubscribes and complaints across Discord and Reddit, compounded by robotic customer support.

Cursor: Composer 1.5 at half price, but platform stability worries

Cursor discounted its Composer 1.5 model 50%, while users report auto-switching to Auto mode, frequent disconnects, slow-queue placement, and opaque Auto billing—some are evaluating alternatives.
• Cursor status update

Arena overhauls image/video leaderboards: scenario-based, noise filtered

Arena clustered 4M user prompts to build scenario-specific leaderboards (commercial design, 3D modeling, etc.), removing ~15% noise. Video Arena also migrated from Discord to the web; Veo 3.1 currently tops it.
• Image Arena update
• Veo 3.1 tops Video Arena

P402.io: auto model-selection and cost accounting layer for OpenRouter

P402 tracks per-model request costs, compares cost-effectiveness (e.g., Opus 4.6 vs Sonnet 4.5), suggests substitutions, and supports USDC/USDT payments with 1% fees.

AuditAI: corrective RAG for NIST CSF compliance auditing

Open-sourced AuditAI uses LangGraph-based Corrective RAG (CRAG) to audit security policies against NIST CSF 2.0, with semantic routing between fast and deep-retrieval paths, mandatory evidence citations, and Llama 3.3 70B + Groq in evaluation.
• Backend code
• Web demo

Industry & Business

Qwen3-Coder-Next praised as "smartest small model" in local circles

Despite the "Coder" name, it performs well as a general-purpose model—chat, recommendations, practical advice—and is considered one of the most reliable all-round small models at its size.
• Reddit review

Is local LLM the next wave? Community runs the GPU-vs-subscription math

Most agree local capability still trails closed cloud models short-term, but small models are improving fast; for privacy-sensitive teams with stable workloads, a $5k–10k local machine increasingly makes sense.
• Reddit discussion

Salesforce executive exodus to OpenAI and AMD fuels strategy speculation

Recent departures include Slack and Tableau CEOs plus the company president and CMO—read as talent voting on growth prospects and possibly signaling a Salesforce AI strategy repositioning.
• Analysis

Cloudflare revenue tops $2B; Vercel CEO rescues customer with $46K bill

After a developer's $46K Vercel bill for rendering Jmail HTML went viral, the Vercel CEO offered to optimize the architecture and cover costs. Cloudflare separately reported over $2B in annual revenue.
• Cloudflare earnings

a16z invests in Japan's Shizuku AI, betting on anime-style AI companions

a16z led a round into Shizuku AI Labs, founded by the creator of a hit AI VTuber, combining Japanese character design with frontier models for long-term companion AI agents.

Policy, Governance & Safety

OpenClaw's architecture makes "indirect-privilege jailbreaks" easier

The BASI community criticizes OpenClaw's loose permission design and weak system prompts, which let agents bypass intended boundaries to access sensitive data. Suggested mitigations include embedding whitelists plus grammatical constraints.
• OpenClaw repo

GPT-5.2 and Opus 4.6 jailbreak arms race: ENI, Glossopetrae evolve

Attackers report partial ENI success on Opus 4.6 and evolving "virtual-world language" bypasses like Glossopetrae; some teams are packaging these as paid red-team services.
• Opus 4.6 ENI jailbreak thread

Discord requires ID uploads from some users; developers push back

Age/ID verification requirements for certain channels and features drew strong developer resistance—some suspect pre-IPO compliance, others fear mass data collection.

OpenAI tests ads in ChatGPT, sparking commercialization debate

Ads will appear in some search/informational scenarios. The community accepts commercialization as inevitable but worries sponsored results and竞价-style ranking could erode trust in model outputs.
• OpenAI ads announcement

KOKKI Agent-Auditor Loop: systematically reducing hallucinations

The proposed KOKKI v15.5 architecture splits models into a "drafting agent" and a "ruthless auditor" enforcing Audit(Draft(input)) on all outputs, recommending cross-model auditing (e.g., GPT drafts, Claude audits) over single-model self-review. Early tests show notable safety/factuality gains.

---

📌 Source: Easy AI Daily (zhichai.net)