JustVugg/colibri is a pure-C, zero-dependency inference engine by Vincenzo Fornaro that streams routed MoE experts from NVMe, running models like GLM-5.2 and Kimi K3 on consumer hardware. 51,687 stars as of 17 November 2026.
Read ArticleExplore our latest tutorials, tech updates, and career advice written by industry experts to help you level up your skills.
JustVugg/colibri is a pure-C, zero-dependency inference engine by Vincenzo Fornaro that streams routed MoE experts from NVMe, running models like GLM-5.2 and Kimi K3 on consumer hardware. 51,687 stars as of 17 November 2026.
Read ArticleWHAM-RT (arXiv:2611.22497) replaces next-token prediction with two-stage discrete diffusion. The paper reports a 79x model speed-up over WHAM, 17 fps, and training from about 28 hours of gameplay.
Read ArticleSeer (arXiv:2611.26423) is an inference-time framework for small deep-research agents. The paper reports GAIA accuracy rising from 13.6% to 41.7% on Qwen3.5-4B and S32 reaching 79.6% on GAIA.
Read ArticleHiSTFormer (arXiv:2611.25610) is a hierarchical spatiotemporal world model. The paper reports it beats strong baselines on MemoryMaze and HM3D within training length and generalises beyond it.
Read ArticleJac, from Prof. Jason Mars (University of Michigan, Jaseci Labs), is a Python-superset language with AI agents as first-class citizens, demonstrated at All Things Open (Raleigh, Oct 2026). Its 1.8–8.7x code-reduction claim needs its caveat.
Read ArticleGoose is a memory-safe systems language by Wouter van Oortmerssen (FlatBuffers, Lobster): no heap allocator, GC, lifetimes or destructors, values inline on compiler-managed data stacks. v2026.3 released 21 Oct 2026.
Read ArticleZero is an experimental systems language from Vercel Labs (Chris Tate & Matt Van Horn), shipped 15 May 2026. It compiles to sub-10 KiB native binaries without LLVM and is designed so AI agents can read compiler output.
Read Articlekarpathy/nanochat is Andrej Karpathy's minimal, hackable Python/PyTorch harness for training a ChatGPT-style model end-to-end on one GPU node. GPT-2-grade model for about $48; 60,705 stars as of 19 November 2026.
Read ArticleFareedKhan-dev/kimi-k3-in-c is a from-scratch portable C99 inference engine (Apache 2.0) running Moonshot AI's Kimi K3 on a single CPU — no GPU, no BLAS, no ML framework. A teaching artifact, not a daily driver.
Read ArticleNVIDIA unveiled the Rubin HPC GPU at SC26 (Chicago, 15–20 Nov 2026) — an HPC-focused chip in the Vera Rubin platform with over 50,000 CUDA cores, 7th-gen NVLink and a claimed 3x Blackwell performance. Availability expected 2028.
Read ArticleParGLU (arXiv:2611.20663) pairs SwiGLU with ReLU² branches in a partially gated FFN. The paper reports it beats SwiGLU at 75M and 152M, with mixed results at 30M.
Read ArticleAnthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on 9 Nov 2026 — up to 30% faster output than Fable 5.1, a 20% price cut, 1M-token context and an optional fast mode up to 2.5x faster.
Read ArticleAsk about courses, website & more
Request human callback:
Please log in to download. You will stay on this page after logging in!
New here? Register now →