Today’s lead · X
dots3-note preview aims at long-horizon open agents, not just bigger models
dots3-note preview is pitched as a 280B MoE model with 16B active parameters, 512K context, multimodal inputs, and Apache 2.0 open weights. The interesting bit for builders is the agency stack: TEMPO training, memory updates, tool use, and new real-life agent benchmarks.
Top signals
4 more
Z.ai’s GLM-5.3 leans into scaled post-training for coding agents
GLM-5.3 is presented as Z.ai’s latest open-source model for complex, long-horizon coding tasks. The claim is open-source SOTA in agentic coding, plus emergent vulnerability discovery and cyber-defense behavior from massive post-training scaling on the same base.
Pika puts generative audio behind one API shape
Pika’s new audio models cover speech, sound effects, soundtrack, and general audio through the Pika API. A builder example says Codex generated music and SFX for a 49-second video through MCP for $0.02; Pika claims pricing up to 20x cheaper than other audio models.
Codex multi-agent orchestration is moving toward model delegation
Codex multi agents v2 now supports delegation to any supported model, including Luna. The practical shift is orchestration: Sol and Terra are described as collaborative peer agents that can message and recursively delegate, while Luna and older models act as leaf agents for narrower delegation.
AMD’s physical AI push moves from keynote theme to embedded hardware stack
AMD is positioning physical AI as a growth market beyond datacenter AI, with Ryzen AI Embedded X100 processors, Kria modules and dev kits, and a broader software-library plan. For robotics and edge builders, the signal is an x86-plus-accelerator stack rather than another server GPU story.
Tools & repos
4 selected
Inferock Bench
A local proxy that records LLM API calls across OpenAI-, Anthropic-, Gemini-, and OpenRouter-shaped requests. Useful if you need per-call usage, failures, retries, and an independent billing receipt instead of trusting vendor dashboards.
MakazhanAlpamys/Soup
Soup promises one-YAML LLM fine-tuning, with layer streaming meant to train an 8B model on a 4 GB laptop GPU. Worth watching if low-VRAM training is your bottleneck.
cursor/plugins
Cursor’s plugin specification and official plugins are now visible as a TypeScript repo. For teams standardizing editor workflows, this is the place to track what Cursor considers extension-compatible.
Attyn
Attyn brings AI actions to the cursor on macOS: inline rewriting, realtime dictation, screen-aware questions, and visual explanations. The useful angle is provider flexibility: credits, your own key, or supported local models.
Blogs worth your time
1 reads
Building an AI Text Detector From Scratch
Raschka turns AI-text detection into a full builder project: dataset construction, DistilBERT classifier training, local API/UI deployment, and using the detector as a verifier. The useful part is the caveat-heavy framing, not detector absolutism.
Community discussions
2 threadsApple Silicon inference is fast hardware waiting on a coherent software stack
A LocalLLaMA deep dive argues Apple Silicon inference is fragmented across mlx-lm, vllm-metal, forks, and conversions. The builder takeaway: prefix caching plus speculative decoding matter, but no Mac stack yet matches CUDA maturity.
A 5090 owner reports 880 tok/s on Qwen3.8-27B with NVFP4 and NInfer
The post reports unusually strong single-RTX-5090 numbers for Qwen3.8-27B on NInfer with NVFP4. The caveats are as important as the speed: Blackwell-only FP4 cores, a non-upstream patch, closed-ish artifacts, and limited validation.
Funding & acquisitions
0 movesNothing material cleared the editorial bar today.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.