If you want to run a model on your own machine today, Ollama and llama.cpp are the safe defaults on a laptop, MLX-LM is worth a look on a Mac, and vLLM or SGLang are the clear picks for a server. When I compared a year of commit history across 12 tools, two once-popular options had effectively stopped: Hugging Face’s TGI is in maintenance mode and its GitHub repository is now archived, and GPT4All hasn’t had a commit since May 2025. If a guide still recommends either for a new setup, it’s out of date.
Live tracker: the numbers in this article are a snapshot from 15 October 2026. For the current status of these tools, and more than 100 open-source AI tools in total, each with a 0 to 100 health score, see the AI Tool Health Index, which refreshes from GitHub automatically. Four of these are also compared side by side: Ollama vs llama.cpp and vLLM vs SGLang. One difference to expect: the tracker counts every commit, including merge commits, while the table below leaves merges out. For projects that merge pull requests with merge commits, such as Open WebUI and Jan, the tracker’s numbers and year-on-year change can look quite different from ours.
The only one of these I’ve run myself is LM Studio, which isn’t in the table because its code isn’t public. On my computer it was slow. That’s common with local models on ordinary hardware, and it’s why I’d match the tool and the model size to your machine before anything else.
Local LLM tools need constant updates: every new model family brings a new architecture, chat template or quantisation format. A runner that stops updating doesn’t break overnight, but it quietly stops supporting the models you’ll want next. That makes maintenance a practical question, not an academic one.
How This Was Measured
Method: on 15 October 2026 we cloned the public GitHub repositories of 12 tools for running language models locally or on your own servers, and compared two identical 91-day windows: 17 July to 15 October 2025, and 17 July to 15 October 2026. For each window we counted commits on the default branch (excluding merge commits), distinct commit-author email addresses (excluding bots), and, for 2026, commits that credit an AI coding tool as co-author. We also read each licence file and README. The scripts were written and run with AI assistance in a cloud sandbox, and the key findings were checked by hand. We did not install or benchmark any of these tools, and closed-source apps such as LM Studio aren’t included because there’s no public history to check.
Commit counts are a rough signal: squash-merging, vendored upstream code and release bots all affect them. The big shifts, and what projects say about themselves, are the useful part.
The Results
| Tool | Best for | Commits 2025 → 2026 | Authors 2025 → 2026 | Last commit | Licence |
|---|---|---|---|---|---|
| llama.cpp | The engine under many local apps | 852 → 1,605 | 187 → 413 | 15 Oct 2026 | MIT |
| Ollama | Easiest local setup | 233 → 339 | 43 → 27 | 15 Oct 2026 | MIT |
| MLX-LM | Apple Silicon Macs | 124 → 130 | 34 → 82 | 15 Oct 2026 | MIT |
| Jan | Desktop chat app | 894 → 930 | 38 → 17 | 15 Oct 2026 | Apache 2.0 |
| KoboldCpp | Desktop app built on llama.cpp | 1,093 → 1,810 | 194 → 405* | 14 Oct 2026 | AGPL-3.0 |
| LocalAI | OpenAI-compatible local server | 492 → 1,477 | 19 → 62 | 15 Oct 2026 | MIT |
| Open WebUI | Web interface for local models | 1,233 → 2,041 | 70 → 51 | 14 Oct 2026 | Open WebUI License (see below) |
| vLLM | GPU serving at scale | 2,708 → 4,300 | 624 → 1,104 | 15 Oct 2026 | Apache 2.0 |
| SGLang | GPU serving at scale | 1,885 → 4,760 | 342 → 729 | 15 Oct 2026 | Apache 2.0 |
| TextGen (formerly text-generation-webui) | Desktop app | 220 → 1 | 9 → 1 | 17 Aug 2026 | AGPL-3.0 |
| TGI (Hugging Face) | GPU serving | 18 → 0 | 7 → 0 | 21 Mar 2026 | Apache 2.0 |
| GPT4All | Desktop chat app | 0 → 0 | 0 → 0 | 27 May 2025 | MIT |
* KoboldCpp regularly merges in llama.cpp’s code, so its author count includes llama.cpp contributors; its own lead developer made 284 of the 2026 commits. LocalAI’s 2026 total includes many automated commits; 408 were made by people.
Commits, a year apart
Non-merge commits on each default branch, 17 Jul to 15 Oct, 2025 vs 2026 (log scale; 0 shown at the left edge)
Source: EvalAITools analysis of public GitHub history, cloned 15 October 2026. GPT4All had no commits in either window and isn’t shown.
Two Tools I’d No Longer Start With
Hugging Face TGI: officially in maintenance mode
TGI’s README says it “is now in maintenance mode”, accepting only minor bug fixes, documentation and light maintenance. Hugging Face itself recommends moving to vLLM, SGLang, or local engines such as llama.cpp and MLX. There were no commits at all in the 2026 window; the last was on 21 March 2026.
Update, 17 October 2026: our live tracker now shows TGI’s GitHub repository as archived, which makes it read-only. Its README still carries the maintenance-mode notice.
GPT4All: no commits since May 2025
The most recent commit in GPT4All’s repository was a CI update on 27 May 2025, nearly 17 months before our check, and the README still advertises support for DeepSeek R1 distillations from early 2025. The README doesn’t say the project has stopped, but nothing has changed, so it won’t pick up newer model architectures.
One to Watch: TextGen
The project long known as text-generation-webui has been renamed TextGen (the old GitHub address redirects). It had 220 commits a year ago and just one in the 2026 window, a README update on 17 August; its development branch was last updated on 16 May 2026. That may be a pause rather than an end, but I’d check before relying on it for new models.
Licences Worth Reading
- Open WebUI uses its own licence: BSD-style terms plus a clause forbidding removal or alteration of the “Open WebUI” branding, unless your deployment has no more than 50 end users in any rolling 30-day period, you have written permission, or you hold an enterprise licence. Contributors also sign a contributor licence agreement. That’s fine for most personal and small-team use, but a company planning a white-labelled internal tool should read it first.
- KoboldCpp and TextGen are AGPL-3.0, which has obligations if you modify them and offer them to others over a network.
- The rest are MIT or Apache 2.0.
Where the Momentum Is
- llama.cpp nearly doubled its commits and more than doubled its contributors (187 to 413). Because Ollama, KoboldCpp, Jan, LM Studio and others build on it, its health matters for the whole local ecosystem.
- vLLM and SGLang are the busiest projects here by far: over 4,000 commits each and more than 700 authors each in 91 days.
- MLX-LM, Apple’s Mac-native runner, kept a steady pace while more than doubling its contributors (34 to 82).
- Ollama grew its commits by 45% while its author count fell (43 to 27). The top contributors are Ollama staff, so this reads as a core team doing more of the work, not a decline.
- Jan also grew its commits slightly while its author count halved (38 to 17); two developers made most of the 2026 commits.
How Much Is AI-Assisted?
vLLM stood out: 1,510 of its 4,300 commits in the 2026 window (35%) credited an AI coding tool. SGLang was at 776 of 4,760 (16%) and MLX-LM at 7 of 130 (5%). Everything else was at 2% or below, including llama.cpp (12 of 1,605) and Ollama (3 of 339). A low share mainly reflects each project’s conventions, since many developers use AI tools without adding a co-author line. Disclosure: Claude, made by Anthropic, helped with this analysis.
What I’d Use
| Situation | My pick | Avoid for new setups |
|---|---|---|
| First local model on a laptop | Ollama | GPT4All |
| Maximum control, newest model support | llama.cpp directly | |
| Apple Silicon Mac | MLX-LM, or Ollama | |
| A desktop chat app | Jan, or KoboldCpp | GPT4All; check TextGen’s status first |
| A browser interface for a team | Open WebUI (read the branding clause) | |
| Serving models on GPUs | vLLM or SGLang | TGI |
| A drop-in OpenAI-compatible API | LocalAI, Ollama or vLLM |
Small models such as MiniCPM5-2B ship guides for several of these runners, so check which ones a model supports before you choose.
FAQ
Is GPT4All still maintained?
Its GitHub repository has had no commits since 27 May 2025. It still works for the models it already supports, but it isn’t being updated for newer ones.
Is Hugging Face TGI deprecated?
Its README says it’s in maintenance mode and recommends vLLM, SGLang, llama.cpp or MLX instead, and as of 17 October 2026 our tracker shows the GitHub repository as archived (read-only).
Ollama or llama.cpp?
Ollama is easier: one install and a simple model library. llama.cpp gives you more control and usually supports new models first, since Ollama and many other tools build on its work.
Is Open WebUI open source?
Its source is public under its own licence, which is BSD-style but forbids removing the Open WebUI branding for deployments over 50 users in 30 days without permission or an enterprise licence.
Related: MiniCPM5-2B for local agents · Which RAG tools are still maintained · Which AI agent frameworks are still maintained
Sources
- GitHub repositories cloned on 15 October 2026: ggml-org/llama.cpp, ollama/ollama, ml-explore/mlx-lm, janhq/jan, LostRuins/koboldcpp, mudler/LocalAI, open-webui/open-webui, vllm-project/vllm, sgl-project/sglang, oobabooga/textgen, huggingface/text-generation-inference, nomic-ai/gpt4all
- TGI README, GPT4All README and Open WebUI LICENSE, as of 15 October 2026
