Cool FOSS projects

gistfile1.txt

https://github.com/xiaoiver/infinite-canvas-tutorial


npx @neuralnomads/codenomad --launch --password dummy
https://github.com/NeuralNomadsAI/CodeNomad

https://github.com/openchamber/openchamber?tab=readme-ov-file

https://github.com/Octane0411/opencode-plugin-openspec

https://www.jsdelivr.com/package/npm/openviking-opencode

https://github.com/srclight/srclight

https://github.com/JuliusBrussee/caveman

Voice to text GUI
https://github.com/jhj0517/Whisper-WebUI

Terminal on webpage
https://github.com/coder/ghostty-web

https://github.com/Fission-AI/OpenSpec/

WebGPU graphics library
https://github.com/q5js/q5.js

LLM inference
https://burn.dev/
https://github.com/jmaczan/torch-webgpu

Color conversion
https://culorijs.org/
https://github.com/color-js/color.js

# Triangulation
https://github.com/mapbox/earcut

# OpåenCode skills for business
https://github.com/sabiertas/business-skills-pack

# Like kubernetes but with untrusted random boxes, kinda
https://github.com/sambigeara/pollen

# Binary app bundle from python that works on all platforms
https://github.com/metaist/cosmofy

# SWE skills
https://github.com/addyosmani/agent-skills

# Markdown LLM based research
https://x.com/karpathy/status/2039805659525644595

# P2P hosting and compute
https://github.com/sambigeara/pollen


For a 32CPU 128Gb RAM server:
numactl --interleave=all ./llama-server \
  -m qwen2.5-moe-35b-q5_k_m.gguf \
  -t 24 \
  -tb 24 \
  --mlock \
  --flash-attn \
  --host 0.0.0.0 \
  --port 8080 \
  --alias gpt-4o
  
  
Visual flowgraph webgl animation editor
https://github.com/idflood/ThreeNodes.js

WebGL resources
https://github.com/sjfricke/awesome-webg
  
Visualize graphs of thousands of nodes and edges using WebGL,
https://github.com/jacomyal/sigma.js

Test in regl/webgl and plotting
https://github.com/gl-vis/regl-scatter2d

Run small LLMs in transformers.js or webllm
https://www.sitepoint.com/webgpu-vs-webasm-transformers-js/
https://llm.mlc.ai/

Time series one-shot forecasting
https://github.com/google-research/timesfm

Hardware Bottleneck: Pure CPU inference is limited by memory bandwidth, not CPU cores. Running giant 70B+ dense models will drop performance to an unusable 1–3 tokens per second (t/s).Model Selection: To match online frontier speed, use a highly optimized Mixture of Experts (MoE) model like Qwen 2.5 35B MoE or Mixtral 8x22B in GGUF (Q4_K_M or Q5_K_M) quantization. This delivers a responsive 15–25 t/s.Parallelism Strategy: Avoid network parallelization across multiple machines over your 1Gbps network; the network latency will stall your inference. Rely entirely on single-machine multi-threaded parallelism.Thread Optimization: Limit compute to 24 threads on your 32-core machine. Saturating all 32 cores causes memory bus fighting and actually slows text generation down.

Enable Transparent HugePages (THP) to minimize memory address lookup overhead.Adjust /etc/security/limits.conf to allow unlimited memlock permissions.Force the CPU governor to performance mode to eliminate clock-speed micro-stutters.Use numactl --interleave=all if your system uses multiple NUMA nodes (e.g., dual sockets, Threadripper, EPYC).