A daily digest of AI industry news for February 25, 2026, originally compiled by Easy AI Daily on zhichai.net.
Key points
Models & Capabilities
- Alibaba Qwen3.5 Medium series: New open models including Qwen3.5-Flash (1M context, built-in tools), 35B-A3B (MoE), 122B-A10B (MoE), and a dense 27B. Community tests suggest the 35B/122B models match or beat the prior 235B generation. GGUF, int4 and lower-bit quantizations, and SGLang support available; aimed at local/edge inference.
- OpenAI GPT-5.3-Codex: Fully available via the Responses API, priced around $1.75 input / $14 output per million tokens (community-reported), positioned as a premium coding/agent model. Official support added for docx/pptx/csv/xlsx files as direct context input.
- Inception Mercury 2: A diffusion-based reasoning LLM claiming ~1000 tokens/s in production. Reviewers note intelligence is not top-tier but coding/agent performance is decent; main selling point is throughput and latency for voice assistants and multi-turn agents.
- Liquid AI LFM2-24B-A2B: 24B-param MoE activating ~2.3B per token, designed to run in 32GB VRAM (~112 tok/s on AMD CPU, ~293 tok/s on H100). Supported in llama.cpp, vLLM, SGLang with GGUF quantizations.
- Qwen3.5 on-device: 4-bit MLX runs ~34–35 tok/s on Mac Studio M3; the MLX build currently lacks vision. Qwen3-VL-2B hooked to a home security camera shows strong scene understanding.
- Claude Code Remote Control: Continue CLI coding sessions from your phone; enterprise Cowork/plugin updates for team workflows.
- Cursor Cloud Agents: Agents run code and tests in cloud environments and return demo videos instead of diffs ("demos, not diffs"); sudo-level operations still limited.
- OpenAI Responses API WebSocket: Reported ~30% speedup in agent workflows; pairs with GPT-5.3-Codex for a full long-context coding/document agent stack.
- AGENTS.md research: LLM-generated long context files reduce success rates and raise costs; hand-written concise notes help slightly. Recommendation: only include key constraints and interface info.
- OpenRouter: New
openrouter/freeauto-routing to free models; GPT-5.3-Codex listed with unified benchmark and pricing display. - Meta × AMD 6GW deal: Meta will deploy ~6GW of AMD Instinct GPUs (mostly MI300X) over five years; engineering blog details the RRCLLX communication library — read as a real challenge to NVIDIA.
- MatX $500M Series B: Building MatX One LLM accelerator with splitable systolic arrays, SRAM for low-latency compute, and HBM for long-context bandwidth. Karpathy commented on compute + two-tier storage as a key bottleneck.
- FlashAttention-3: Official prebuilt PyTorch wheels for CUDA 12.6+/13, x86/ARM, Linux/Windows; LibTorch ABI compatible, Python≥3.10, Torch≥2.9.
- llama.cpp regression: Latest master builds fail to load Qwen3.5 GGUF ("Failed to read magic"); workaround is rolling back to build 8145 or using a stable release.
- GPU MODE: eBPF extended to GPU devices/drivers, PyTorch Helion 0.3.0 (adaptive tuning, Triton-TileIR bridge), and NVSHMEM/CUDA debugging discussions.
- Capability–reliability gap (Princeton et al.): Reliability decomposed into 12 dimensions across model generations — capabilities surge but reliability improves only marginally; agents still fail in long-tail scenarios (compared to autonomous driving's edge-case problem).
- OpenClaw security study: Splitting dangerous instructions into innocuous steps bypasses current guardrails; an open-source fix proposed — multi-step tool calls need holistic risk analysis.
- SWE-bench Multilingual: 300 tasks across 9 languages, decontaminated from SWE-bench Verified; best models solve ~72%, and language rankings invert across models.
- OCR benchmarks saturating: Many models score 90%+ on OmniDocBench yet fail on real complex PDFs; text-extraction-then-LLM often beats end-to-end image pipelines for PDF QA.
- Nature optimizer paper questioned: Researchers suspect poorly tuned baselines (or test-set hyperparameter selection) inflate the claimed wins; calls for reproduction against public nanogpt baselines.
- Claude Code COBOL demo → IBM stock drop: An Anthropic blog post on analyzing/modernizing COBOL systems was read as threatening IBM consulting; IBM fell over 10% intraday, though the demo still requires heavy human review.
- Perplexity/Comet voice mode: Upgraded voice mode for all users; some Pro users complain about tighter rate limits favoring enterprise/Max tiers.
- Edge vision case study: 0.7GB Qwen3-VL-2B-Instruct (IQ2 quant) on a MacBook M3 + SharpAI Aegis narrates security-camera footage in natural language.
- Small tools: OpenClaw repurposed for printer-maintenance and watch-deal bots; Manus.im users request unlimited-chat plans; Kollect turns forms into real-time AI conversations.
- Gemini app updates: A user built a Metal Gear-style mini game with Gemini 3.1 Pro in ~2 hours; the mobile app adds video templates (Veo 3.1 quality criticized but barrier to entry lowered).
- Anthropic "distillation attacks": Anthropic alleges DeepSeek, Moonshot, and MiniMax used ~24k fake accounts for 16M+ Claude calls to distill its models, arguing safety properties don't transfer. Community pushback: everyone trains on scraped data, and some see the narrative as justification for tighter China export controls.
- Anthropic's open-source record: Critics note Anthropic has never open-sourced an LLM or tokenizer, sparking safety-vs-openness debates.
- Distillation vs copyright: Discussions compare distilling from closed APIs with training on copyrighted web data; both occupy legal gray zones.
- Jailbreak evolution: Communities iterate prompt techniques (persona dialogue, chain-of-thought conflicts) against Kimi 2.5, Gemini 3.1 low, and DeepSeek.
- SWE-bench Verified retired: OpenAI deprecated the benchmark, citing memorization of original commits among frontier models and flawed remaining tasks.