Notes from Sai Dutta Abhishek Dash — AI infrastructure & security engineer — on LLM gateways, code-security agents, self-hosted systems, and the open-source tools he ships.
IFM released six foundation models from 0.9B to 375B, plus roughly 21.5TB of the actual training datasets, all Apache 2.0. I verified the dataset sizes and benchmarked the three small models myself: the 0.9B does real reasoning at 19 tokens per second in 2.2GB of VRAM. Here is what shipped, what is still missing, and where the catches are.
I stress-tested Ternary Bonsai 2 27B on two Tesla T4s: 15 tokens per second on a single weak GPU, the full 262k context loading in 26GB of VRAM, and a complete todo webapp built and self-tested inside a local coding agent. The claims mostly check out, with a few catches.
I benchmarked LFM2.5 350M and Qwen3.5-2B on my iPhone SE 2 (2020) with 3GB of RAM using the PocketPal app. One model hits 70 tokens per second. The other needs 22 minutes for a single pass. Here is where the line is.
Two developments are breaking NVIDIA's CUDA monopoly: the Vulkan compute backend for llama.cpp (cross-vendor LLM inference on any GPU) and the Mojo programming language (a CUDA alternative that compiles to any accelerator).
Cloudflare Workers AI serves 50+ models from its edge network with 10,000 free Neurons a day. Since July 2026 the frontier coding models need the $5/mo Paid plan, but the free tier still covers real daily coding through OpenCode. What is free, what moved to Paid, and how I wire it up.
India hit E20 five years early, capped blending at 20% in July 2026, and legally approved E100 - while 300 million older vehicles run E20 with no opt-out. Fresh September 2026 numbers on CIAN Agro, OMC losses, ethanol prices, and the flex-fuel rollout.
Two stories keep getting merged into one: the Bartz v. Anthropic copyright case ($1.5B settlement) and Anthropic's claim that Moonshot AI routed 300,000 customer requests through Claude to harvest responses. The public record supports a narrower conclusion than either side's loudest claims.
Real Qwen3.8-27B numbers on the DGX Spark GB10: the memory wall caps single-stream decode at 11-12 tok/s, and speculative decoding is the only way out. SGLang with DSpark hits 34 tok/s, veloGB10 reaches about 40 tok/s, and NVFP4 keeps quality within noise of FP8.
I ranked nine free GPU notebook platforms against Google Colab in 2026. Kaggle (Tesla P100, 30 hrs/week, 29GB RAM) and AWS SageMaker Studio Lab (T4, 4 GPU hrs/day, persistent storage) take S tier, Paperspace Gradient is A tier with persistent storage, and Colab itself lands B tier for vague limits and 2-3 hour disconnects.
