Key points
Models & Capabilities
- Alibaba releases Qwen3.5 Medium series: Open-source mid-size models including Qwen3.5-Flash (1M context, built-in tools), 35B-A3B MoE, 122B-A10B MoE, and a 27B dense model. Community reports the 35B/122B models approach or exceed the prior 235B generation. GGUF, low-bit quantizations (int4 and below), and SGLang support are ready, targeting local/edge inference. Official release | 35B-A3B on HuggingFace | 122B-A10B on HuggingFace
- OpenAI launches GPT-5.3-Codex: Fully available via the Responses API at roughly $1.75 input / $14 output per M tokens, positioned as a premium coding/agent model. New file input support for docx/pptx/csv/xlsx. Developer announcement
- Inception Mercury 2: A "reasoning diffusion LLM" claiming ~1000 tokens/s in production. Reviews say intelligence is not top-tier but coding/agent performance is decent; the selling point is throughput and latency. Artificial Analysis review
- Liquid AI LFM2-24B-A2B: 24B MoE with ~2.3B active parameters per token, designed for 32GB VRAM. Reported speeds: ~112 tok/s on AMD CPU, ~293 tok/s on H100. Supports llama.cpp, vLLM, SGLang, with multiple GGUF quants. Blog
- Qwen3.5 on-device experiences: 4-bit runs at ~34–35 tok/s on Mac Studio M3; the MLX build currently lacks vision. Qwen3-VL-2B has been attached to home cameras with good scene understanding. Reddit thread
- Claude Code Remote Control: Continue terminal coding sessions seamlessly on your phone; enterprise Cowork/plugin updates for team-wide workflows. Announcement
- Cursor Cloud Agents: Agents run your programs in the cloud, run tests, and return demo videos instead of diffs ("demos, not diffs"). Free cloud environments; sudo-level operations still restricted. Intro page
- Responses API WebSocket: Real-world agent workflows reportedly ~30% faster. Benchmark
- AGENTS.md study: Auto-generated long context files often reduce success rates and raise costs; short hand-written notes help slightly but add tokens. Keep them minimal. Paper summary
- OpenRouter: New
openrouter/freeauto-routing to free models; GPT-5.3-Codex listed with unified benchmark scores and effective pricing. Rankings - Meta × AMD 6GW deal: ~6GW of AMD Instinct GPU deployment (mostly MI300X) over five years, alongside Meta's custom RRCLLX communication library — a challenge to NVIDIA's dominance. Engineering blog
- MatX raises $500M Series B for MatX One, an LLM accelerator combining SRAM low-latency compute with HBM bandwidth. Karpathy notes compute + two-tier memory scheduling as a key bottleneck. Karpathy comment
- FlashAttention-3 official wheels: Prebuilt for CUDA 12.6+/13, x86/ARM, Linux/Windows, LibTorch ABI compatible; pip-installable with Python ≥3.10, Torch ≥2.9. Download page
- llama.cpp regression: Latest master builds fail to load Qwen3.5 GGUF with "Failed to read magic"; workaround is rolling back to build 8145 or using an official release.
- GPU MODE highlights: eBPF extended to GPU contexts (arXiv:2512.12615), PyTorch Helion 0.3.0 (release), NVSHMEM performance exploration.
- Capability–reliability gap (Princeton et al.): Reliability split into 12 dimensions; capabilities surge while reliability improves only marginally — agents still fail in long-tail scenarios. Paper thread
- OpenClaw security study: Splitting dangerous instructions into innocuous sub-steps bypasses existing agent guardrails; an open-source fix is proposed.
- SWE-bench Multilingual: 300 tasks across 9 languages, best models solve ~72%; language rankings can invert. Announcement
- OmniDocBench saturation: Many models score 90%+ but real-world complex PDFs still fail; extracting text before PDF QA often beats end-to-end image input.
- Nature optimizer paper questioned: Reviewers suspect poorly tuned baselines or test-set hyperparameter selection; calls for reproduction with public nanogpt baselines.
- Claude Code COBOL blog post moves IBM stock: Anthropic's demo of modernizing COBOL systems (bank mainframes, ATM cores) was read as an automated migration threat; IBM fell over 10% intraday — though it's a capability showcase, not a turnkey tool. Reddit discussion
- Perplexity/Comet new voice mode for all users; some Pro users complain about tighter rate limits.
- Qwen3-VL-2B security camera: A ~0.7GB IQ2 quant runs on MacBook M3 + SharpAI Aegis, producing narrative descriptions like "a courier approaches the door and drops a package." Reddit case
- Gemini updates: A user built a simple Metal Gear-style game in ~2 hours with Gemini 3.1 Pro; the mobile app adds video templates. 9to5Google
- Anthropic "distillation attacks": Anthropic claims DeepSeek, Moonshot, and MiniMax used 24,000 fake accounts to make 16M+ calls against Claude for distillation training, arguing safety properties don't transfer. Community debates whether this is also a narrative push for stricter China-related compute export controls. Official blog
- Distillation vs. copyright: Parallel discussions ask whether training on closed-API outputs differs fundamentally from training on copyrighted web data.
- SWE-bench Verified deprecated: OpenAI retires it, citing memorization of original commits by frontier models and flaws in remaining unsolved tasks. Announcement
Agents & Tooling
Infrastructure & Hardware
Research & Methods
Products & Applications
Industry, Policy & Safety
📌 Source: Easy AI Daily