GitHub Pulse
The sum of commits in the most recent complete week across a curated watchlist of major open-source AI/ML repositories, read directly from each repo’s own weekly commit-activity statistics on GitHub. The delta is the week-over-week change in that same figure.
Comparing the week of 2026-10-18 (5381 commits) against the week of 2026-10-11 (4153 commits), across 16 of the 20 repos on the watchlist that had data available at computation time.
Per-repo breakdown
Every repo on the watchlist that GitHub returned data for, sorted by this week’s commit count.
| Repository | This week | Prior week | Change |
|---|---|---|---|
| deepseek-ai/deepseek-harness | 1332 | 528 | +152.3% |
| unslothai/unsloth | 1190 | 793 | +50.1% |
| vllm-project/vllm | 1150 | 1007 | +14.2% |
| QwenLM/qwen-code | 659 | 524 | +25.8% |
| open-webui/open-webui | 332 | 306 | +8.5% |
| pytorch/pytorch | 231 | 482 | -52.1% |
| ggml-org/llama.cpp | 172 | 190 | -9.5% |
| huggingface/transformers | 153 | 130 | +17.7% |
| anthropics/anthropic-sdk-python | 54 | 27 | +100.0% |
| langchain-ai/langchain | 49 | 29 | +69.0% |
| ollama/ollama | 29 | 45 | -35.6% |
| huggingface/diffusers | 18 | 19 | -5.3% |
| ggml-org/whisper.cpp | 6 | 61 | -90.2% |
| mistralai/mistral-common | 6 | 5 | +20.0% |
| run-llama/llama_index | 0 | 6 | -100.0% |
| meta-llama/PurpleLlama | 0 | 1 | -100.0% |
Day by day
The most recent days this number was computed. Not every calendar day has a row — a day with no row simply wasn’t computed that day.
| Date | Value | Change |
|---|---|---|
| Oct 29, 2026 | 5,381 | -2.0% |
| Oct 28, 2026 | 5,490 | +7.7% |
| Oct 27, 2026 | 5,097 | — |
Why these repos, and why commits
The watchlist is roughly 20 major open-source AI/ML projects spanning inference engines (llama.cpp, vLLM), core frameworks (PyTorch, Transformers, Diffusers), application layers (LangChain, LlamaIndex, AutoGen, Open WebUI, ComfyUI), and several labs’ own SDKs and model repos. Commit volume is a real, if noisy, proxy for how much active engineering is happening across the ecosystem day to day — it’s naturally choppier week to week than a usage metric like downloads or tokens, so a single week’s swing (up or down) shouldn’t be read as a trend on its own.
A repo drops out of the table above only when GitHub’s commit-activity endpoint didn’t return usable data for it at computation time (rate-limited, or GitHub still computing its stats cache) — it’s excluded from both weeks’ totals for that run, not counted as zero.
Computed daily from GitHub’s own public commit-activity statistics, with no AI model involved in the calculation. See the full Signal methodology for how every other number on the Signal page is calculated.