Easy AI Daily News Digest - February 17, 2026

A full roundup of the day's AI industry developments, curated from the Easy AI daily report on zhichai.net.

Models and Capabilities

Alibaba Releases Qwen3.5-397B-A17B: Open-Source 400B-Class Multimodal MoE

Alibaba Cloud open-sourced Qwen3.5-397B-A17B, combining hybrid linear attention with sparse MoE: 397B total parameters, 17B activated, support for 201 languages, native 256K context expandable to roughly 1M, under the Apache-2.0 license. vLLM and other frameworks offer day-zero support; community analysis estimated KV cache overhead at ~31KB/token, keeping long-context inference viable in BF16/FP8. The Qwen3.5-Plus API offers 1M context with search and a code interpreter, though some users criticize pricing.

Links: Official release | Architecture/KV discussion | vLLM deployment | Qwen3.5 blog

MiniMax M2.5 and GLM-5: New Chinese "Holiday Models"

MiniMax M2.5 has 230B parameters with 10B active and 200K context, benchmarked at ~2500 tok/s/GPU on 8x H200 with vLLM. Token-level process rewards reportedly improve RL signal efficiency. Local deployment needs ~200GB VRAM; 2x RTX 6000 Blackwell reaches 120-130 tok/s. GLM-5 is praised for tool calling and multi-turn agent tasks, though service stability is still maturing.

Links: MiniMax technical post | SemiAnalysis benchmarks | Unsloth local guide

Anthropic Opus 4.6: 1M Context + Automatic Self-Checking

Claude Opus 4.6 supports 1M-token context and appends a "check your work" step that can overturn earlier errors. LMArena testers found it retains long-session context better. Strict hourly rate limits remain for heavy users.

Step 3.5 Flash: Cheap but Punchy Reasoning Model

OpenRouter community comparisons find Step 3.5 Flash dramatically outperforms its price class, though platform support and routing lag behind major closed models. Comparison video

CommonLID: New 109-Language Language-ID Benchmark

From Common Crawl and EleutherAI, CommonLID shows top LangID models score below 80% F1 even on languages they claim to support, revealing that prior benchmarks were too easy for real web data. Paper | Dataset

Agents and Tooling

OpenAI Acquires OpenClaw

OpenClaw creator Peter Steinberger joins OpenAI to work on personal agents; OpenClaw moves to a foundation and stays open source. The community celebrates a solo-dev-plus-agents success story while criticizing its rough config, hidden costs (idle heartbeats burn money), and fearing a "ClosedClaw" future.

Links: Altman confirmation | Reddit skepticism | steipete's technical writeup

"Harness Engineering" as the New Moat

Consensus is emerging that agent experience depends less on the model than on the harness: tool orchestration, context management, lifecycle, evaluation, and observability. Minimal alternatives (PicoClaw, nanobot) are appearing; LangChain/LangSmith pitch "traces as the agent's stack trace."

OpenClaw in the Wild

Users granted it root SSH to upgrade Proxmox from v6 to v8 automatically; others built "agent companies" with task-decomposing lead bots and frontend/backend bots, Tavus video-call integration, and SEO pipelines mass-producing hundreds of articles.

MCP Spec Debate: The "Token Tax" of Structured Output

Developers complain APIs lacking native schema support force JSON Schema into prompts, wasting tokens. Proposals: distinguish text/image/object result types, put schemas in metadata, and pass timezone/user context explicitly rather than storing server-side state.

Jazz: Terminal-Based Agent CLI

Open-source Jazz reads code, runs git, calls MCP services, sends email, writes release notes, and supports scheduled tasks. Related: Crowdcent wraps DSPy as an MCP server; Cloudflare experiments with Markdown for agents.

Infrastructure and Hardware

• NVIDIA GB300 NVL72 and the power bottleneck: ~50x perf/MW and 35x lower cost/token vs. Hopper are claimed, but the industry bottleneck has shifted from GPUs and HBM to datacenter power and distribution. Western Digital's 2026 HDD capacity is reportedly sold out, with some AI customers locked through 2027/2028.
• AccelOpt: agents writing CUDA kernels: Self-optimizing LLM agents reportedly beat FlashInfer 0.5.3 by 1.5x on GQA paged decode and 1.38x on prefill. 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 mismatches plague B200 users.
• WebGPU + BitNet: BitNet-B1.58 2B runs at ~125 tok/s on an M4 Max via the open-source Hesper library.

Research and Methods

• CoVe, RLM, and Rubric RL: Meta's Chain-of-Verification roughly doubles accuracy via answer-then-verify prompting; Omar Khattab's Recursive Language Models favor recursive code-calling over longer attention; Cameron Wolfe surveys 15+ rubric-based RL papers replacing fuzzy LLM judges.
• Model genealogy: Matrix-driven weight homology (arXiv:2508.06309), black-box independence tests reconstructing Llama fine-tune trees (arXiv:2502.12292), and black-box provenance (arXiv:2510.19796) could help detect rebranded models.
• Assistant Axis: A new paper shows activations drift along a persona "assistant axis" during long conversations — measurable evidence of long-chat behavior drift.
• Safety research: X-Ware's diffusion-based activation edits surface "meta-neurons" beyond SAEs; FAR.AI warns that training against deception probes may teach models to hide internal states rather than become honest.
• QED-Nano 4B: Lewis Tunstall's IMO-focused math model uses multi-stage distillation and inference caching to spend compute on hard problems.

Products and Applications

• Perplexity backlash: Deep search quotas cut from 200 to 20/month; matching old usage would cost ~$167/month vs. $20; TrustPilot down to 1.5/5 as users migrate to Claude/Opus 4.6 or Kimi.
• Kimi & MiniMax reliability issues: Kimi K2.5 often impresses (with a $40/month API plan fitting OpenClaw), but CLI install failures, duplicate billing, quota bugs, and scam clone sites push users toward self-hosting large MoEs on 700GB RAM + 200GB VRAM rigs.
• Coding workflows: Practitioners pair Claude Cowork for pipeline tasks with planners like Ergo/planbot and executors (Codex, Claude Code, OpenClaw). The lesson: the winning stack is "planning + version control + observability," not any single model.
• Security applications: PassLLM fine-tunes Qwen3-4B LoRA on millions of leaked passwords for personal-info-based guessing; ATIC uses three Claude Opus 4.5 instances to score uncertainty and escalate to humans.

Industry and Policy

• Chinese "Spring Festival model wave": Qwen3.5, GLM-5, MiniMax 2.5, and ByteDance's Seedance 2.0 (with a Jia Zhangke-directed demo short) launched in quick succession; video generation is seen as entering real director workflows.
• Anthropic vs. Pentagon: Per Axios, the DoD may label Anthropic a "supply-chain risk" because it restricts use for mass surveillance of Americans and fully autonomous weapons.
• Gemini distillation attack: Google says attackers sent 100K+ prompts attempting to clone Gemini without source or weights; the community notes irony given Google's own web-scraping training data. Ars Technica
• ChatGPT Lockdown Mode: OpenAI adds enterprise tool-call restrictions (caching, reduced web access) to cut prompt-injection and data-exfiltration risk.
• Reproducibility debate: Criticism of an OpenAI physics paper using GPT-5.2 without disclosing prompts, tools, or model details; calls for journals to require conversation logs.

---

📌 Source: Easy AI Daily (zhichai.net)