Easy AI Daily News — February 17, 2026
A roundup of AI industry developments for February 17, 2026, covering model releases, agent tooling, infrastructure, research, products, and policy.
Models and Capabilities
Alibaba releases Qwen3.5-397B-A17B: open-source 400B-class multimodal MoE
Alibaba Cloud launched the open-source Qwen3.5-397B-A17B, combining hybrid linear attention with sparse MoE: 397B total parameters with 17B activated. It supports 201 languages, ships with native 256K context extendable to ~1M, and is released under Apache-2.0. Frameworks like vLLM offered day-zero support; community analysis estimated KV cache overhead at ~31KB/token, and long-context inference runs in BF16/FP8. The API version, Qwen3.5-Plus, offers 1M context with search and a code interpreter, though the community complains pricing is high.> Links: Official release tweet | Architecture and KV estimation discussion | vLLM support and deployment recipe | Qwen3.5 blog (with spatial intelligence examples)
MiniMax M2.5 and GLM-5: a new batch of Chinese "holiday models"
MiniMax released M2.5 with 230B parameters, 10B activated, and 200K context, measured at roughly 2500 tok/s per GPU on 8×H200 + vLLM. The company emphasizes token-level process rewards for better RL signal efficiency. Community testing indicates ~200GB VRAM for local deployment; 2× RTX 6000 Blackwell reaches 120–130 tok/s. GLM-5 is described as "smart and conversational," strong at tool calling and multi-turn agent tasks, though service stability and routing are still being tuned.> Links: MiniMax technical post | SemiAnalysis performance test | Unsloth local-run guide
Anthropic Opus 4.6: 1M context plus automatic self-checks
Claude Opus 4.6 launched with 1M-token context and a "check your work" self-verification step at the end of answers that can overturn earlier mistakes. LMArena users testing with large code instructions found it remembers earlier context better in long sessions and corrects errors at the end. However, Anthropic still enforces strict hourly rate limits for Opus 4.6.> Links: LMArena community testing
Step 3.5 Flash: a cheap but capable reasoning model
OpenRouter community comparisons found Step 3.5 Flash offers exceptional price-performance, outperforming its price tier on many benchmarks and real tasks. Platform support remains limited and routing/deployment lag behind better-known closed models.> Links: Performance comparison video
CommonLID: a new language-ID benchmark across 109 languages
Common Crawl, EleutherAI and others released CommonLID, a language identification benchmark for multilingual web text covering 109 languages. Testing shows top LangID models score under 80% F1 even on languages they claim to support, suggesting prior benchmarks were too easy and the real web is much harder.> Links: CommonLID paper | Hugging Face dataset
Agents and Tooling
OpenClaw acquired by OpenAI: personal agents vs. open-source shells
OpenClaw author Peter Steinberger joined OpenAI to work on personal agents; OpenClaw itself moves to a foundation to remain open source. The community celebrates it as a "one person + agents built an acquisition-grade product" story, while critics note it mainly chains APIs/MCP together, is rough to configure, has hidden costs (a 30-minute heartbeat burns money), and fear a post-acquisition "ClosedClaw" future.> Links: Altman acquisition confirmation | OpenClaw usage and skepticism (Reddit) | steipete's technical deep-dive on OpenClaw
"Harness engineering" becomes the new moat: agents are more than models
Multiple discussions argue the deciding factor in agent experience isn't the base model but the surrounding harness: tool orchestration, context management, lifecycle, evaluation, and observability. Good harnesses compound, and vendors use interaction data for RL. Minimal alternatives like PicoClaw and nanobot are emerging, while LangChain/LangSmith pushes observability-first with "traces as the agent's stack trace."> Links: Ben Burtenshaw on harnesses | LangSmith tracing and debugging
OpenClaw in the wild: from Proxmox upgrades to multi-agent teams
OpenClaw serves as a high-stakes automation testbed: one user gave it root SSH access to upgrade a home Proxmox cluster from v6 to v8 including reboots and error handling. Others built an "agent company" where a tech-lead bot decomposes tasks for frontend/backend bots, integrated Tavus for video calls with facial expressions and screen sharing, and ran SEO pipelines producing hundreds of articles.> Links: #showcase collection
MCP spec discussion: the "token tax" of structured output
In the MCP Discord, developers complain many APIs lack native schema support, forcing JSON Schema into prompts as pseudo structured output and wasting tokens. Suggestions include clearly separating text/image/object result types and placing object schemas in metadata, plus passing context like timezones explicitly as parameters rather than storing state server-side.> Links: MCP discussion thread
Jazz and other terminal agents: MCP, Git, and Shell in one CLI
The open-source Jazz project builds a terminal-resident agent that reads code, runs git, calls MCP services, sends email, writes release notes, and supports scheduled jobs. Related work: Crowdcent is packaging DSPy as an MCP server, and Cloudflare is experimenting with HTTP endpoints returning Markdown for easier agent consumption.> Links: Jazz project | Cloudflare: Markdown for agents
Infrastructure and Hardware
NVIDIA GB300 NVL72 and the power bottleneck
Analysis claims GB300 NVL72 delivers ~50× per-MW performance improvement over Hopper and 35× lower cost per token, but the industry bottleneck has shifted from GPU counts and HBM supply to datacenter power and distribution. Western Digital's 2026 HDD capacity is reportedly fully booked by enterprise customers, with some AI customers locked through 2027/2028.> Links: NVL72 metrics discussion | AI-driven HDD demand | Infrastructure bottleneck evolution
FlashInfer contest and AccelOpt: agents writing CUDA kernels
The GPU MODE community is running a B200 kernel optimization contest on FlashInfer-bench, allowing FP8 and other tricks within accuracy tolerances. The AccelOpt team used self-improving LLM agents to modify kernels, claiming 1.5× speedup over FlashInfer 0.5.3 on GQA paged decode and 1.38× on prefill, with code open-sourced.> Links: AccelOpt repo | FlashInfer contest details
GPU performance tuning realities: benchmark jitter and occupancy misreads
Kernel tuners on H100/H200 report simple matmul TFLOPs fluctuating between 1400–1500 in NVProf, and note Achieved Occupancy doesn't count idle SMs — SM active cycles are needed to estimate true active SMs. Meanwhile, the CUTLASS/CuteDSL/Proton stack frequently breaks on B200 due to version mismatches, forcing users into TTGIR docs, layout algebra, and old commits.> Links: Cutlass/CuteDSL issues | Proton tutorial example
WebGPU + BitNet: 2B model at 125 tok/s on M4 Max
A developer ran the BitNet-B1.58 2B model on WebGPU at ~125 tok/s on M4 Max, open-sourcing the Hesper library. Combined with PyTorch on WebGPU, running small models in the browser is becoming practical.> Links: Hesper library
Research and Methods
CoVe, RLM, and Rubric RL: a new trio for long reasoning
Meta's Chain-of-Verification (CoVe) uses answer-then-self-check prompting, summarized as nearly doubling accuracy on many tasks. Omar Khattab's Recursive Language Models (RLM) favor recursive reasoning via "writing code that calls itself" over ever-longer attention. Cameron Wolfe surveyed 15+ rubric-based RL papers that replace vague "LLM-as-judge" scoring with structured rubrics.> Links: CoVe explainer | RLM design discussion | Rubric RL summary
Language model genealogy: from weight lineage to black-box provenance
Eleuther community papers on "model family trees": one reconstructs kinship between LLM weights via matrix homology analysis; another uses independence tests to reconstruct the fine-tuning tree of Llama-family models from black-box access; further work traces provenance via "rewrite memory." Given large lineages like Qwen/Llama, such techniques may become tools against rebranded "wrapper" models.> Links: Matrix-Driven Weight Homology | Independence Tests for Language Models | Blackbox Model Provenance
Assistant Axis: measurable evidence that long chats drift
A new paper extracts activation directions corresponding to personas/roles, defines an "Assistant Axis," and shows activations drift along it in long conversations, shifting behavior from helpful assistant toward something else. Previously subjective user complaints now have measurable evidence.> Links: Steering LLMs by Persona Directions
X-Ware meta-neurons and FAR.AI: reflections on anti-deception training
X-Ware proposes a diffusion model trained on internal activations to generate "activation edits" for cleaner behavioral control, discovering meta-neurons at a higher level than SAEs. FAR.AI warns that naively training against "deception probes" may teach models to hide internal states (activation-level cloaking) rather than become genuinely honest — a trap for future safety evaluations.> Links: X-Ware meta-neurons | FAR.AI deception research
QED-Nano 4B and small math-reasoning models
Lewis Tunstall released QED-Nano 4B targeting IMO-level mathematical proofs. It uses multi-stage distillation and inference caching so small models can "expand aggressively" at inference, spending compute on hard problems rather than everyday chat — useful for local math/theorem-proving setups.> Links: QED-Nano introduction
Products and Applications
Perplexity slashes quotas and raises prices; users flee
Perplexity Pro cut deep searches from 200 to 20 per month, added upload limits and 7-day data retention. Long-time users calculate that maintaining previous usage would cost ~$167/month (vs. $20 before); TrustPilot rating dropped to 1.5/5. Many also feel models became "dumber with worse memory" since early February and are migrating to Claude / Opus 4.6 or Kimi.> Links: Perplexity Discord discussion
Kimi & MiniMax: unstable APIs and billing issues make local LLMs look better
Kimi K2.5 performs well — sometimes better than Sonnet/Opus 4.5 at coding and reasoning — with a $40/month API plan fitting OpenClaw. But complaints persist: CLI plugin install failures, subscription disappearance with repeated charges, jumping quotas, plus scam clone sites. Many are seriously considering self-hosting these large MoEs with 700GB RAM + 200GB VRAM local setups.> Links: Kimi docs and CLI guide | Kimi Discord billing feedback
Claude Code, Claude Cowork, and OpenClaw: real-world coding workflows
In Latent Space and elsewhere, practitioners are settling on patterns: Claude Cowork for pipeline tasks (e.g., auto-uploading Zoom recordings to YouTube), planners like Ergo/planbot to decompose features, then delegating execution to Codex/Claude Code/OpenClaw. The takeaway: what works is not one model but the full "planning + version control + observation" workflow.> Links: Ergo planner skill examples | Claude Cowork workflow discussion
Security applications: PassLLM password auditing and ATIC uncertainty scoring
PassLLM fine-tunes Qwen3-4B LoRA on millions of real password pairs for personal-info-based password guessing, with alarming demo accuracy. ATIC uses three independent Claude Opus 4.5 instances in a "three-brain" setup, scoring each answer for aleatoric uncertainty and knowledge gaps, escalating to humans or specialized systems above thresholds — a nascent "reliability score" module pattern for LLMs.> Links: PassLLM project | ATIC system
Industry and Company News
China's "Spring Festival model wave": Qwen3.5, GLM-5, MiniMax 2.5, Seedance 2.0
This Spring Festival was jokingly called "model release week": beyond Qwen3.5, GLM-5 and MiniMax 2.5 launched, plus ByteDance's Seed/Seedance 2.0 on video — even commissioning director Jia Zhangke for a demo short. The community feels video generation has moved from toy to real director workflows, with China's investment in humanoid robots and video models seen as "producing real results, not show homes."> Links: Model roundup | Seedance 2.0 cinematic discussion
OpenClaw author, after two Anthropic legal warnings, joins OpenAI
Memes circulated: Anthropic twice sent legal warnings to OpenClaw over its name, but the author was ultimately hired by OpenAI and the project acquired — "Claws 🦞 biting back at Anthropic." Commentators contrast the companies: Anthropic leans toward models and safety research, OpenAI toward products and growth.> Links: Meme discussion (Reddit)
Perplexity, Stripe, Apple: platform fees and AI investment pacing
Latent Space founders complain Stripe's add-ons effectively take 8.3% of revenue, with European card fees still calculated at 2.9% — deemed excessive. Others speculate Apple is deliberately waiting out the current AI spending wave, planning to deploy cash for partnerships or acquisitions once training/inference marginal costs drop, avoiding the current "$2 trillion capex sprint."> Links: Stripe fee complaints | BuccoCapital on AI capex
Policy, Governance, and Safety
Pentagon threatens to label Anthropic a "supply chain risk" over usage terms
Axios reports the US Department of Defense is considering labeling Anthropic a "supply chain risk" over usage restrictions, pressuring contractors to drop Claude. Anthropic refuses to allow the model for mass surveillance of Americans and fully autonomous weapons, insisting on "reasonable scenarios" only, while the Pentagon insists all legal uses must be available. Many drew parallels to PRISM, raising the real question of whether safety terms can cost vendors military contracts.> Links: Axios report | Reddit discussion 1 | Reddit discussion 2
Google report: attackers made 100,000+ prompts trying to "distill" a Gemini clone
Google says attackers queried Gemini over 100,000 times attempting model distillation to clone a cheaper version without touching source code or training data. Google framed it as "IP theft," but the community questions how effective such distillation could be, and jokes that Google's own web-scraping training practices make the IP argument awkward.> Links: Ars Technica report
OpenAI launches ChatGPT "Lockdown Mode": trading capability for control
OpenAI added Lockdown Mode for enterprise/business ChatGPT, restricting or rewriting tool calls (cached search, weakened web access) to reduce prompt-injection and data-exfiltration risk. It's an acknowledgment that tool-using LLMs are a new attack surface, and some companies prefer predictable, compliant agent behavior over capability.> Links: Feature discussion
Models in science without disclosure? OpenAI physics paper sparks reproducibility debate
Researchers questioned an OpenAI physics paper using GPT-5.2 without releasing prompts, tool chains, or model details — yet the journal accepted it. Black-box models in papers make results hard to reproduce; critics urge journals to require conversation logs or tool configs. Authors and OpenAI executives later offered clarifications, but skepticism remains.> Links: Critique thread | OpenAI-side response
---
📌 Source: Easy AI Daily