Easy AI Daily News | 2026-02-18
A roundup of AI industry news for February 18, 2026, translated and summarized from the original Chinese forum post.
Models and Capabilities
Anthropic launches Claude Sonnet 4.6: 1M context, near-Opus quality, new cost/quality tradeoffs
Anthropic released Sonnet 4.6, billed as the "most capable Sonnet," with upgrades across coding, computer use, long-context reasoning, and agent planning, plus 1M token context (beta) at the same price as 4.5. It scored ELO 1633 (first place) on third-party GDPval-AA, but consumed roughly 4.8x more tokens than Sonnet 4.5 (also more than Opus 4.6), so total cost on complex tasks may actually be higher. Official benchmarks: 79.6% SWE-bench Verified, 58.3% ARC-AGI-2. Users preferred it over Opus 4.5 in 59% of internal preference tests. Early reports of function-name hallucinations and degraded structured output were quickly fixed.
Links: Official release | Benchmark data | Artificial Analysis on token usage | Reddit discussion
Qwen3.5-397B-A17B: Alibaba's open-source 397B MoE with multimodal, million-token context
Alibaba released Qwen3.5-397B-A17B: a 397B-total / 17B-active MoE supporting native 262K context (expandable to 1M), 200+ languages, and image/video multimodality. The company claims near-GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro performance on instruction following, multilingual knowledge, and video understanding. Community GGUF quantizations exist, but VRAM requirements make it mostly a cloud/cluster model.
Links: Official blog | Hugging Face | Unsloth GGUF | Reddit discussion
GLM-5 technical report: DSA, async RL, and agent training
Zhipu published the GLM-5 technical report (arXiv:2602.15763) and demo. Community tests show strong performance on benchmarks like WeirdML, sometimes matching or beating closed models. The report highlights DSA architecture, asynchronous RL infrastructure, and agent-scenario RL algorithms. Some developers reported underwhelming results in certain IDE plugins, possibly due to integration and inference stack issues.
Links: arXiv:2602.15763 | Community benchmarks
Tiny Aya: Cohere's 3.35B multilingual model that runs on phones
Cohere Labs released Tiny Aya, a 3.35B-parameter open model covering 70+ languages, designed for phones/edge devices and reportedly trained on just 64 GPUs. The goal is a practically deployable multilingual assistant rather than top benchmark scores.
Links: Announcement
Agents and Tooling
OpenClaw goes viral — and gets users banned
OpenClaw, a multi-agent framework for controlling real computers, spread rapidly, with users building project management, CRM, and browser automation. But users connecting via Claude OAuth reported account bans, citing Anthropic's terms against reverse-engineered third-party access. Security concerns abound: giving LLMs read/write access to local machines, browsers, and keys is risky if prompt injection or malicious skills slip through.
Links: OpenClaw showcase | ClawRecipes
LangChain pushes "Harness Engineering": turning production logs into agent testbeds
LangChain advocates mining online traces to build automated evaluations and self-verification loops for agent workflows, alongside TerminalBench and scheduled LangSmith Insights evaluations.
Links: LangSmith Insights | OpenAI's blog post
MCP discusses paid tools: payment signals in the protocol
MCP community proposal SEP-2007 would let MCP servers signal "payment required" (starting with X402), enabling agents to call paid tools within budgets. Maintainers are cautious, preferring URL-redirect approaches for now; not a near-term priority.
Links: PR #2007
DirectShell: the accessibility layer as a universal app interface
DirectShell controls apps via OS accessibility APIs instead of screenshots and coordinate clicks, abstracting the accessibility layer into a unified DOM/control tree — making any windowed software scriptable like the web.
Infrastructure and Hardware
GPU MODE community: squeezing TFLOPS from B200 / RTX cards
Engineers shared custom matmul and FlashInfer MoE tuning details: ~350→368 TFLOPS on hand-written persistent CUDA kernels, 5.7x speedup from FlashInfer's MoE baseline on B200, plus FP8/FP4 numerical error discussions and Nsight Compute / CUDA events usage tips.
Links: matmul_V3 source | FlashInfer baseline
CoDA-GQA-L: KV cache from 160GB to 136MB
From the Eleuther community, CoDA-GQA-L fixes each layer's KV cache at 384 slots (256 recent tokens, 64 landmark tokens, 64 exponential-moving-average summaries), compressing KV usage from 160GB to 136MB on long sequences — practical for 1M-context applications like search, long-form writing, and long-task agents.
Links: Paper (Zenodo) | Code
MXFP4 / NVFP4: FP4 really needs Blackwell
Unsloth and Modular both confirm FP4's advantages only appear on Blackwell Tensor Cores (compute capability ≥ 12.0); on Ampere and older cards, simulation often makes it slower than FP8. Buying FP4-capable setups on older cards is likely wasted money.
Links: Unsloth docs | Modular forum
Cloud inference outages: OpenRouter/HF/Perplexity 401s and 500s
OpenRouter suffered widespread 401s (war-room investigation); Perplexity API users hit 401s despite valid balances; Hugging Face Inference Endpoint users saw 500 "service panicked" errors despite all-green status pages, requiring endpoint rebuilds. Practical lesson: keep emergency scripts to fully rebuild endpoints.
Links: OpenRouter status | Hugging Face status
Research and Methods
- Agent World Model: 1,000 executable environments, 35,062 tools, 10,000 tasks for RL-training tool-using agents. Overview
- Lossless Context Management (LCM/Volt): hierarchical DAG compression with recoverable references; reportedly more stable than Claude Code from 32K to 1M context on OOLONG. Details
- Moltbook: a 2.6M-agent social simulation (300K posts, 1.8M comments) found macro-level culture converges and stabilizes, but individual agent "influencer" effects are statistical noise. Discussion
- Every Eval Ever: the EvalEval alliance proposes a unified BIDS-like data/metadata standard for LLM benchmarks, prompted by issues like Cybench's non-randomized flags inflating scores. Project | Cybench
- Preventative steering: adding interference vectors in representation space to test whether models hold their goals; possibly usable as representation-space data augmentation. arXiv:2602.15029
- SpeakType: fully local, open-source Mac dictation app based on Whisper. GitHub | Site
- Manus Agents: a personal agent with long-term memory and tool integrations (Gmail, Notion) accessible via Telegram. Announcement
- Food Truck Bench: 12 LLMs each got $2,000 and a food truck for a 30-day simulation; only 4 survived. Claude Opus 4.6 earned ~$49K, GPT-5.2 ~$28K. Nearly all models that took loans went bankrupt. Reddit | Vending-Bench 2
- Mistral acquires Koyeb: folding the serverless platform and team into Mistral Compute, moving from model vendor to full-stack compute provider. Announcement
- Grok 4.20 criticized: community found the model ID is still grok-4-1-thinking-1129, mocking it as "four Grok 4.1 agents in a trench coat"; also criticized for political bias echoing Musk's positions and weaker performance vs. peers. Discussion
- Anthropic allows military use with two red lines: no mass surveillance, no autonomous weapons. Discussion
- Waymo economics: sixth-gen robotaxi platform costs ~$70K per vehicle, projected to halve by 2028; 500K+ driverless rides per week, ~3x annual growth. Analysis
- OpenClaw OAuth bans: Anthropic treats third-party desktop-control OAuth access as a TOS violation. Lesson for desktop-agent teams: use official APIs or self-hosted models, not vendor web OAuth. Discussion
- Dark web intel from BASI community: GitLab projects with maintainer access auctioned ($200 starting, $2,000 flash sale); "unlimited AI" services on Tor — LLM tech and traditional cybercrime are now mixing in the same circles. Post
- API key leaks: an OpenRouter user's key leaked despite gitignore and was drained of $10 in 20 minutes. Best practices remain: least privilege, regular rotation, repo-wide key scans before deployment. Discussion
- Cybench flag flaw: early versions used fixed flags from well-known CTFs, letting models recall them from training data; randomizing flags sharply dropped success rates — a reminder that sloppy eval design inflates security benchmark numbers. Cybench
Products and Applications
Sonnet 4.6 rolls out across dev tools
Sonnet 4.6 launched day-one in Cursor, Windsurf, Cline, Perplexity, and OpenRouter. Cursor says it's clearly better than 4.5 on long tasks but slightly below Opus 4.6 in "intelligence"; Cline offered a free trial (70% of developers preferred 4.6 in internal A/B tests); Perplexity integrated it into Deep Research and browser agents. Trend: Sonnet 4.6 as the cost-effective workhorse, Opus for maximum quality.
Links: Cursor | Windsurf | Perplexity
PolyAI raises $200M, launches Studio Lite
Voice-AI company PolyAI raised $200M (with Nvidia, Khosla participating) and launched Agent Studio Lite, which builds a working voice agent from a website URL in 5 minutes; 3-month free trial for brand customers.
Links: Announcement
Other launches
Industry and Business
Policy, Governance, and Security
📌 Source: Easy AI Daily (zhichai.net translation)