Today’s lead · x.com
draft launches a local team knowledge base for coding agents
draft is a desktop app and CLI that creates a separate local context folder for team agents, syncing installed skills, MCP servers, and shared context without mixing it into working files or code. It says it can continuously update that context from Granola meetings, Slack channels, and agent sessions, and works across Claude Code, Codex, Cursor Agent, Hermes, and OpenClaw.
Top signals
17 morevLLM shares a four-node DGX Spark recipe for Nemotron 3 Ultra 550B
vLLM highlighted NVIDIA’s guide for serving Nemotron 3 Ultra 550B-A55B-NVFP4 across four DGX Spark machines as one OpenAI-compatible endpoint. The setup uses tensor parallelism size 4, vLLM’s nightly aarch64 OpenAI container, RoCE networking, a 262,144-token max context, Mamba cache flags, and optional MTP speculative decoding.
Parallel Code runs Claude Code, Codex, Gemini, Copilot, and Antigravity in isolated worktrees
Parallel Code is a free MIT-licensed macOS/Linux app for dispatching multiple coding-agent tasks in parallel, with each task getting its own git branch and worktree. Its linked docs describe diff review, inline comments, phone monitoring over Wi-Fi or Tailscale, Docker sandboxing via project-specific Dockerfiles, CI status watching, and support for Claude Code, Codex CLI, Gemini CLI, Copilot CLI, and Antigravity CLI.
squad gives Claude, Gemini, Codex, and OpenCode a shared SQLite message bus
squad is a Rust CLI that lets multiple AI CLI agents collaborate through one-shot shell commands backed by SQLite, with no daemon or background service. It provides roles such as manager, worker, and inspector, supports slash-command setup for Claude Code, Gemini CLI, Codex CLI, and OpenCode, and can optionally launch tiled tmux sessions with isolated git worktrees.
OpenRouter publishes continuous GPQA and TAU-Bench results for open-weight models
OpenRouter says it continuously runs GPQA and TAU-Bench on most open-weight models and uses those results in AutoExacto, its default meta-benchmark for routing tool calls. The linked GLM 5.2 page describes a 1M-token context model from Z.ai with high and xhigh reasoning efforts, OpenAI-compatible API access, rolling 30-day provider price, latency, TTFT, throughput, success-rate, and eval charts.
GLM-5.2 beats Claude Code on Semgrep’s IDOR vulnerability benchmark in reported independent eval
A report says Z.ai’s GLM-5.2, released June 16 under an MIT license, scored about 39% F1 on Semgrep’s IDOR vulnerability detection benchmark versus Claude Code’s 32%. The same post says GLM-5.2 identified vulnerabilities at roughly one-sixth the cost and is downloadable from Hugging Face for standard open inference stacks.
Codex desktop app gets a “Build iOS Apps” plugin with simulator and Xcode MCP support
A post says OpenAI’s Codex desktop app added a “Build iOS Apps” plugin that gives Codex skills for iOS Simulator, Liquid Glass UI, interface elements, Xcode Build MCP, local debugging, and testing. The described workflow keeps code local: the user describes an app, Codex writes native iOS code, launches the simulator, and iterates with visible source.
vLLM-Omni details TTS serving optimizations for Qwen3-TTS, VoxCPM2, Higgs Audio V3, and Fish Speech S2 Pro
vLLM-Omni published an engineering deep dive on why TTS serving needs different scheduling than text LLMs, including first-audio latency budgets of a few hundred milliseconds and continuous audio across streaming chunks. The team reports +61.5% audio throughput and nearly halved P99 latency for Qwen3-TTS on H20×2, +172% throughput for VoxCPM2, a 2.7x Higgs Audio V3 speedup, and a q_len=1 Triton attention kernel for Fish Speech S2 Pro.
Baidu Unlimited-OCR now runs in vLLM with constant KV cache
vLLM added support for Baidu’s Unlimited-OCR document parser, using Reference Sliding Window Attention to keep KV cache fixed during long decoding. The recipe claims one-shot transcription of 40+ pages under a 32K context budget, 35% faster output than DeepSeek-OCR at 6K tokens, and BF16 inference on a single GPU with at least 8GB VRAM via a dedicated vLLM Docker image.
CHORD uses contact-wrench guidance to transfer human dexterous manipulation demos to robots
CHORD is a robot-learning framework that represents demonstrations by the forces and torques contacts induce on objects, rather than only where contact occurs. The project reports 82.12% average success across 1,831 benchmark tasks, 90.77% whole-body manipulation success, a 4,739-task sim-ready dexterous benchmark, and transfer to real dexterous hands, with code planned for the Video to Data repo.
Beads passes 650K downloads as a persistent memory layer for coding agents
Beads is a distributed graph issue tracker that gives coding agents structured persistent memory using a Dolt-backed database, dependency graphs, JSON output, and auto-ready task detection. Its docs describe integrations for Codex, Claude Code, Factory Droid, Cursor, and other agents, plus commands such as bd ready, bd update --claim, bd remember, and bd prime.
MiniMax M3 runs locally across three Macs with Parallax and GenericAgent
Gradient says it ran MiniMax M3, a 428B model, locally on three Macs using Parallax, then had GenericAgent create a five-stock portfolio and write it to disk. The demo emphasizes local execution with no cloud calls or API bills, using a roughly 3,000-line self-evolving agent on desk-side hardware.
Nous Hermes Agent dashboard shown running a 35B MoE locally on DGX Spark
A post describes NousResearch’s Hermes Agent dashboard as a local web UI for chat, sessions, files, models, skills, logs, and channels. The setup shown runs a 35B mixture-of-experts model called Ornith in near-lossless FP8 through vLLM on a single DGX Spark, with tokens streaming in the browser while GPU logs run in the terminal.
Community DSpark and DeepSeek V4 implementation reaches 60 tok/s on dual GB10 hardware
A community repo is described as a first DSpark plus DSV4 implementation using a purpose-built vLLM fork. The post says the setup was run day zero on a dual NVIDIA GB10 system and confirmed 60 tokens per second decode.
TPU architecture writeup explains systolic arrays, large on-chip memory, and pod-scale design
A linked engineering writeup walks through TPUv4 internals, including two TensorCores per chip, 128×128 MXU systolic arrays, VPUs, 32MiB VMEM, 10MiB SMEM, 128MiB CMEM, and 32GiB HBM. It contrasts TPU design with NVIDIA GPUs and cites TPU v5p at 500 TFLOPs per chip, 4.45 ExaFLOPs per 8,960-chip pod, and TPUv7 Ironwood claims of up to 42.5 ExaFLOPs per 9,216-chip pod.
MegaTrain claims full-precision 100B+ LLM training on one GPU by treating VRAM as a cache
A thread introduces MegaTrain as a training system that inverts the memory hierarchy so GPU VRAM behaves like a transient, stateless cache rather than the place the whole model must live. The stated goal is full-precision training of 100B+ LLMs on a single GPU, with the model state streamed through the device instead of resident in GPU memory.
Cerebras founder details the three-year cooling buildout behind wafer-scale AI hardware
Cerebras co-founder Andrew Feldman says the company spent 2016 to 2019 building its own heat sink and water-cooling system because no supplier had components for a wafer-scale chip about 58 times larger than the largest GPU and drawing 15KW. He says the result was one of the first production AI systems to use water cooling and remains necessary for running Cerebras’ wafer-scale engine cooler and more reliably than GPUs.
Hugging Face highlights new papers on agent memory, world models, and real-time streaming
Hugging Face Daily Papers’ weekly list focuses on agentic AI, world models, and streaming models, including Alibaba’s Qwen-AgentWorld and Wan-Streamer v0.1. Other highlighted papers include PlanBench-XL for long-horizon LLM tool-use planning, EnterpriseClawBench for benchmarking agents from real workplace sessions, OpenRath for runtime state in agent systems, and MemSlides for personalized slide-generation agents.
Tools & repos
5 selectedRobbyant/lingbot-map
A feed-forward 3D foundation model for streaming scene reconstruction is relevant to robotics, spatial AI, and real-time perception work.
commaai/openpilot
openpilot remains a major open robotics and driver-assistance codebase for engineers studying production autonomy systems.
xbtlin/ai-berkshire
This Claude Code and Codex-based multi-agent research framework is a useful example of domain-specific agent workflows and adversarial analysis.
simplex-chat/simplex-chat
SimpleX is a substantial private messaging stack for engineers interested in identifier-free communication architectures.
ripienaar/free-for-dev
This long-running free-tier catalog remains a practical resource for developers and founders assembling low-cost infrastructure.
Blogs worth your time
2 readsHP Inc. launches Frontier strategic partnership with OpenAI
From openai.com.
The Thermodynamic AI Chip · Thomas Ahle
From youtube.com.
Funding & acquisitions
5 movesPersona.js
WebMCP-native AI chat for frontends could help developers embed agentic interfaces directly into apps.
Dotient
A local semantic search app is relevant for builders exploring private, on-device knowledge retrieval workflows.
Lyto
A cross-browser, tools, and messages AI agent points toward more unified personal automation surfaces.
discode.ai
A single interface for 100+ AI models may be useful for teams comparing model behavior and switching providers quickly.
GetCompress
Lossless media compression without context switching can streamline asset workflows for product and web teams.
Bengaluru radar
3 events
Red Hat PyTorch x Hugging Face Bengaluru Meetup
A technical meetup covering SGLang, TRL reinforcement learning, PyTorch profiling, distributed training, and GPU-to-GPU communication.
Bengaluru Startup Meetup
A founder and operator meetup focused on early wins, go-to-market challenges, team building, fundraising, and startup growth.
Startup Stories by Bengaluru Founders
A Bengaluru founder community session built around startup journeys, lessons from challenges, and networking with people building companies.