Easy AI Daily | February 25, 2026

Key points

• Alibaba released the Qwen3.5 Medium series: Qwen3.5-Flash (1M context with built-in tools), 35B-A3B (MoE), 122B-A10B (MoE), and 27B (dense). Community testing suggests 35B/122B match or exceed the previous 235B generation. GGUF, int4 and low-bit quantizations, and SGLang support are available, targeting local/edge inference.
• OpenAI launched GPT-5.3-Codex via the Responses API, priced around $1.75 input / $14 output (community-reported), positioned as a premium coding/agent model. Direct file inputs now support docx/pptx/csv/xlsx, and WebSocket support in the Responses API reportedly speeds agent workflows by ~30%.
• Inception Labs released Mercury 2, a diffusion-based LLM claiming ~1000 tokens/s in production—prioritizing throughput and latency over peak intelligence.
• Liquid AI shipped LFM2-24B-A2B: 24B-parameter MoE activating ~2.3B per token, running on 32GB cards (~112 tok/s on AMD CPU, ~293 tok/s on H100); supports llama.cpp, vLLM, SGLang.
• Meta × AMD deal: Meta will deploy ~6GW of AMD Instinct GPUs (mainly MI300X) over five years, with Meta's RRCLLX communication library for large-scale AMD GPU interconnect—seen as a challenge to NVIDIA's dominance.
• MatX raised $500M Series B for MatX One, an LLM accelerator using a partitionable systolic array with SRAM (low latency) plus HBM (long-context bandwidth). Karpathy commented on compute + two-level memory scheduling as a key bottleneck.

Agents & tooling

• Claude Code Remote Control: continue a terminal session from your phone; enterprise Cowork/plugin updates for team-wide custom workflows.
• Cursor Cloud Agents: agents run programs in the cloud, execute tests, and return demo videos instead of diffs ("demos, not diffs"); sudo-level operations still limited.
• AGENTS.md study: a new paper finds auto-generated long context files often reduce agent success rates and raise costs; concise, human-written constraint notes help slightly but still add token cost.
• OpenRouter launched an openrouter/free auto-routing option and added GPT-5.3-Codex, with unified benchmark scores and effective pricing on model pages.

Infrastructure

• FlashAttention-3 official prebuilt wheels for CUDA 12.6+/13, x86/ARM, Linux/Windows; pip-installable with Python ≥3.10 and Torch ≥2.9.
• llama.cpp regression: latest master fails to load Qwen3.5 GGUF ("Failed to read magic", no VRAM allocation); workaround: roll back to build 8145 or use a stable release.
• GPU MODE highlights: eBPF extended to GPU devices/drivers, Helion 0.3.0 with adaptive tuning and Triton-TileIR bridging, NVSHMEM performance exploration.

Research

• Capability–reliability gap (Princeton et al.): reliability measured across 12 dimensions improves only marginally while capabilities surge; agents still fail often in long-tail scenarios. An online dashboard accompanies the paper.
• OpenClaw security study: splitting dangerous instructions into innocuous sub-steps bypasses existing agent guardrails; open-source fix proposed.
• SWE-bench Multilingual: 300 tasks across 9 languages (avoiding SWE-bench Verified data); best models solve ~72%, with language rankings reversing across models.
• OmniDocBench saturation: many models exceed 90%, but real-world complex PDFs still fail; text extraction before PDF QA often beats end-to-end image input.
• Nature optimizer paper challenged: researchers suspect untuned baselines and possible test-set hyperparameter selection; calls for reproduction with public baselines like nanogpt.
• SWE-bench Verified deprecated by OpenAI due to severe contamination—frontier models recall original commits, and many remaining tasks are flawed.

Products & industry

• Anthropic's COBOL blog post: a Claude Code demo on analyzing/modernizing COBOL systems was read as a threat to IBM's consulting business; IBM stock dropped over 10% in a day, though the post was a capability showcase.
• Anthropic distillation accusations: Anthropic claims DeepSeek, Moonshot, and MiniMax used 24,000 fake accounts for 16M+ Claude calls to distill its models. Community debate ensued—some see it as narrative-building for stricter China chip export controls; others compare distillation to training on copyrighted public data. Critics also note Anthropic has never open-sourced any LLM.
• Perplexity/Comet launched upgraded voice modes for all users; some Pro users criticize tighter rate limits favoring Max/enterprise tiers.
• Edge vision: a developer connected ~0.7GB Qwen3-VL-2B-Instruct (IQ2 quantized) to a home camera on a MacBook M3 + SharpAI Aegis, producing full text narratives of scenes (e.g., courier dropping a package).
• Gemini updates: a user built a simple Metal Gear Solid-style game in ~2 hours with Gemini 3.1 Pro; the mobile app added video templates for social media content.

📌 Source: Easy AI Daily