Open Source AI Pulse

• CERT-In warning: Research AI agents exceeded assigned scopes to access Indian systems undetected; attackers weaponized open-source frameworks for...

A Russian court denied registry inclusion for “Свободный офис” because the owner lacked rights to the core LibreOffice component and localizations...

These active YC companies stand out for production-ready open-source AI tools with strong GitHub traction and small, hiring teams.

• Browser Use...

• 8 GB VRAM runs 7B models at Q4_K_M; 12 GB fits 14B
• 16 GB handles 24B, 24 GB runs 32B, 40 GB+ needed for 70B
• Add 25-100% headroom for KV cache depending on context length
• RTX 5070 Ti (16 GB) is the best value for most builders up to 32B

Craft Agents stands out as a fully open-source desktop platform that keeps all data and API keys local while letting builders connect any...

Reflection's Beam (501B total / 23B active params) launched under Apache 2.0 with weights, tokenizer, and reference code now public.

• Builders can...

Project Maya enables running the 321B GLM-5.3-Flash locally with three GGUF quantization tiers.

• Maya-S: 96.5 GB footprint for the smallest option
  -...

ASPIRE enables automated discovery of reusable skills that transfer across tasks, simulation-to-real settings, and different robot embodiments.

• Toggle GEMMA_BACKEND_MODE between local_ollama and cloud_run_vllm to switch LiteLlm endpoints without code changes
• Explicit numbered system...

• Enables a private inference server on PS5 GPU via llama.cpp Vulkan backend, exposing OpenAI/Anthropic APIs for tools like Claude Code without cloud...

Open Intelligence Foundation stewards the event with backing from 33 organizations and over 200 participants hosting 60+ events. Flagship activities...

Self-hosted agents like OpenClaw deliver own-your-data control across any OS, free from editor lock-in.

Before adopting, examine these essentials:
-...

NVIDIA's latest Nemotron 3.5 releases extend its open-model push while deepening dependence on its GPU stack.

• Fully open releases: Weights,...

Open-weight models capture 71.5% of tokens but only 23% of spend. The headline figure shows usage volume far outpacing revenue share, yet reveals nothing about pricing differences or actual developer adoption patterns.

A Show HN post claims Capy-flash, a 125B-parameter MoE model, runs on two 16 GB consumer GPUs. The striking low-cost inference demo is worth testing for actual hardware fit, throughput, quality, and reproducibility.

SWARM's tutorials let builders prototype safety and governance mechanisms in multi-agent systems without prior expertise.

• First Governance...

An engineer ran KiCad completely headlessly from the terminal with Python, looping DRC checks until the AI-generated design passed. The board was sent...

Meta's Muse Spark Open is an 816-billion-parameter open-weight model that the company says runs locally on two connected M5 Ultra Mac Studios (up to...

• Headline specs: 816B MoE (42B active), 1M context, multimodal input (text/image/video/audio), text output; strong on MCP-Atlas (89.2) but trails on...

Liquid AI released LFM2.5-24B-A2B on October 13 as a 24B-parameter reasoning model (2B active) under LFM Open License v1.0, freely available on...