Easy AI Daily Digest (Feb 17, 2026): Qwen3.5, MiniMax M2.5, Opus 4.6, and More

Forum topic · 小凯 · 2026-03-27

Summary

A daily roundup of AI industry news for February 17, 2026. Key items: Alibaba released the open-source Qwen3.5-397B-A17B, a multimodal sparse MoE with hybrid linear attention, 17B active parameters, 201 languages, native 256K context (extensible to ~1M) under Apache-2.0; MiniMax launched M2.5 (230B/10B active, 200K context, ~2500 tok/s per GPU on 8×H200); Claude Opus 4.6 added 1M-token context with self-checking; Step 3.5 Flash drew praise for cost-effective reasoning. The OpenClaw project was acquired by OpenAI and its creator hired, sparking debate on agent 'harness engineering' as a moat. Infrastructure news highlighted NVIDIA GB300 NVL72 efficiency gains shifting bottlenecks to power, plus community kernel-optimization work and WebGPU BitNet inference at ~125 tok/s on M4 Max. Research covered Chain-of-Verification, Recursive Language Models, rubric-based RL, model lineage detection, and an 'Assistant Axis' paper quantifying persona drift in long chats. Policy items included the Pentagon reportedly threatening to label Anthropic a supply-chain risk, Google alleging 100K+ prompts used to clone Gemini, and OpenAI's new ChatGPT Lockdown Mode.

📅 AI Industry Digest — February 17, 2026

This is a structured English summary of the Easy AI daily roundup (originally published in Chinese).

Key points

Models & Capabilities

Alibaba Qwen3.5-397B-A17B: Open-source sparse MoE with hybrid linear attention, 397B total / 17B active parameters, 201 languages, native 256K context (scalable to ~1M), Apache-2.0. Day-zero vLLM support; community estimated KV cache overhead at ~31KB/token. The API version (Qwen3.5-Plus) offers 1M context with search and code interpreter, though pricing drew criticism.
Official tweet | KV cache discussion | vLLM support | Blog
MiniMax M2.5: 230B params / 10B active, 200K context, ~2500 tok/s per GPU on 8×H200 with vLLM. Uses token-level process rewards for RL. Local deployment needs ~200GB VRAM (2× RTX 6000 Blackwell reaches 120–130 tok/s).
Tech post | SemiAnalysis benchmarks | Unsloth guide
GLM-5: Praised for tool calling and multi-turn agent tasks; service stability still maturing.
Claude Opus 4.6: 1M-token context plus a "check your work" self-review step that can override earlier errors; strict hourly rate limits remain. (LMArena discussion)
Step 3.5 Flash: Considered extremely cost-effective on OpenRouter, but platform support lags. (Video comparison)
CommonLID: New 109-language web language-ID benchmark; top models score under 80% F1 even on languages they claim to support. (Paper | Dataset)

Agents & Tooling

OpenClaw acquired by OpenAI: Creator Peter Steinberger joins OpenAI for personal-agent work; OpenClaw moves to a foundation as open source. Community praised the solo-builder achievement but noted fragile config, hidden costs, and "ClosedClaw" concerns. (Altman confirmation | Reddit | Steinberger's writeup)
"Harness engineering" as the new moat: Agent UX depends on orchestration, context management, and observability more than the base model; minimal alternatives (PicoClaw, nanobot) emerging; LangChain/LangSmith push tracing-first debugging. (Thread | LangSmith)
Real-world OpenClaw use cases: Root-SSH Proxmox v6→v8 upgrades, multi-agent "software company" setups with Tavus video calls, and bulk SEO content pipelines.
MCP spec debate: Structured-output "token tax" from stuffing JSON schemas into prompts; proposal to distinguish text/image/object outputs with schemas in metadata, and pass timezone/environment explicitly as parameters.
Jazz terminal agent: CLI agent reading code, running git, calling MCP services, scheduling tasks. (GitHub) Related: Crowdcent's DSPy-as-MCP-Server; Cloudflare's Markdown for agents.

Infrastructure & Hardware

NVIDIA GB300 NVL72: Claimed ~50× performance-per-MW and 35× lower cost-per-token vs Hopper; industry bottleneck shifting from GPUs/HBM to data-center power and distribution. Western Digital's 2026 HDD capacity reportedly booked out, with AI customers locking supply through 2027/2028. (Metrics | HDD demand)
Agent-written CUDA kernels: GPU MODE B200 FlashInfer-bench competition; AccelOpt's self-improving LLM agents claim 1.5× faster GQA paged decode and 1.38× prefill vs FlashInfer 0.5.3. (Repo)
Kernel tuning pitfalls: NVProf TFLOPs jitter (1400–1500) on simple matmuls; Achieved Occupancy excludes idle SMs — use SM active cycles instead; CUTLASS/CuteDSL/Proton version-mismatch errors on B200.
WebGPU + BitNet: BitNet-B1.58 2B at ~125 tok/s on M4 Max via the open-source Hesper library. (GitHub)

Research & Methods

Long-reasoning toolkit: Meta's Chain-of-Verification (answer-then-verify prompting, up to ~2× accuracy); Omar Khattab's Recursive Language Models (recursive code-calling instead of longer attention); Cameron Wolfe's survey of 15+ rubric-based RL papers replacing fuzzy LLM-judge scoring.
Model lineage forensics: Matrix-driven weight homology reconstructs LLM family trees; independence tests recover fine-tuning trees from black-box Llama access; black-box provenance via "rewrite memory" — potential tools against rebranded "shell" models. (2508.06309 | 2502.12292 | 2510.19796)
Assistant Axis: Measurable activation drift along persona directions over long conversations — the first quantitative evidence that assistants "drift" in extended chats. (2601.10387)
Safety notes: X-Ware's diffusion-based activation edits and meta-neurons; FAR.AI warns training against deception probes may teach activation-level camouflage rather than honesty.
QED-Nano 4B: Lewis Tunstall's 4B model for IMO-level proofs via multi-stage distillation and inference caching.

Products & Industry

Perplexity Pro backlash: Deep search quota cut from 200 to 20/month plus upload limits; maintaining old usage would cost ~$167/mo (previously $20); TrustPilot down to 1.5/5; users migrating to Claude/Opus 4.6 or Kimi.
Kimi K2.5: Strong coding/reasoning at $40/mo API tier, but billing glitches, CLI plugin issues, and scam clone sites push users toward self-hosting large MoEs (700GB RAM + 200GB VRAM).
Coding workflows: Consensus that planners (Ergo/planbot) plus executors (Claude Code, Codex, OpenClaw) with version control and observability beat any single model. (Ergo)
Security applications: PassLLM fine-tunes Qwen3-4B LoRA on real password leaks for personalized password guessing; ATIC uses three independent Claude Opus 4.5 instances to score aleatoric uncertainty and knowledge gaps, escalating to humans. (PassLLM | ATIC)
Chinese "Spring Festival model wave": Qwen3.5, GLM-5, MiniMax 2.5, and ByteDance Seedance 2.0 (video, with a Jia Zhangke demo short) all launched around the holiday; video generation seen moving toward real director workflows.
Business chatter: Stripe's effective ~8.3% take criticized; speculation that Apple is deliberately waiting out the AI capex surge before spending its cash.

Policy, Governance & Safety

Pentagon vs. Anthropic: Per Axios, the DoD may label Anthropic a "supply chain risk" over its refusal to allow mass surveillance of Americans and fully autonomous weapons, echoing PRISM-era debates. (Report)
Gemini cloning attempt: Google says attackers made 100,000+ prompts trying to distill a cheap Gemini clone; community both doubts distillation efficacy and notes irony given Google's training-data practices. (Ars Technica)
ChatGPT Lockdown Mode: OpenAI adds restricted tool-calling mode for enterprise/business to reduce prompt-injection and data-exfiltration risk — trading capability for predictability.
Reproducibility dispute: GPT-5.2's uncredited-detail involvement in a physics paper reignites calls for journals to mandate prompts/tool configs or transcripts for model-assisted research.

---

📌 Source: Easy AI Daily