Abandoned Open Source AI Projects and What to Use Instead in 2026

This is a field guide to abandoned open source AI projects and the alternatives worth moving to in late 2026, grouped by category: video animation, text to speech, LLM inference, 3D generation, detection and depth, and vision language models. It is for developers who still have a 2023 or 2024 tutorial, a pinned requirements.txt, or a production pipeline pointing at one of these repos. Every status below comes from each project's GitHub metadata, README, and package index entries, checked in the first week of October 2026.

Keep one rule in mind: inactive does not mean broken. Several of these projects still run, some are downloaded millions of times a month through maintained wrappers, and weights do not rot. What changes is who fixes things when your CUDA, Python, or serving stack moves on.

Abandoned open source AI projects: alternatives at a glance

Last commit means the default branch, not GitHub's updated date, which moves with stars and other repo activity.

Video animation: AnimateDiff and what replaced it

AnimateDiff, the ICLR 2024 spotlight paper that added a plug-in motion module to Stable Diffusion 1.5, has over 12,000 GitHub stars and a long tail of tutorials. The repo was never archived, but its last commit landed on July 17, 2024, and it carries 318 open issues and pull requests. Its requirements.txt pins torch 2.3.1, diffusers 0.11.1, and gradio 3.36.1. For scale, diffusers 0.41.0 shipped on October 6, 2026.

The fair part: AnimateDiff the technique is far from dead. Diffusers 0.41 still ships AnimateDiff pipelines for text-to-video, ControlNet, SDXL, SparseCtrl, and video-to-video, and the ComfyUI-AnimateDiff-Evolved node pack reached version 1.6.0 in July 2026. For short stylized loops on SD 1.5 checkpoints and LoRAs, those wrappers still do the job; just skip the original repo.

For new work, the active options are video models trained for motion from the start:

• Wan 2.2 is Apache 2.0. Its TI2V-5B model generates 720p at 24 fps and, per the README, runs on consumer cards like the RTX 4090. Wan2.2-Animate-14B (September 2025) handles character animation and replacement.
• LTX-2.5 from Lightricks generates synchronized audio and video. The LTX-2 repo shipped v1.4.2 on October 2, 2026. Weights are gated on Hugging Face under the LTX-2.x Community License, and the README's quick-start download of the distilled model plus encoders is roughly 66 GiB.
• Wan2GP is the low-VRAM front end. It runs Wan 2.1 and 2.2, LTX-2 through 2.5, and HunyuanVideo 1 and 1.5, with select models working on as little as 6 GB of VRAM. Version 17.00 landed on October 6, 2026.

HunyuanVideo 1.5 is another open-weight option, but its Tencent Hunyuan Community License states that it does not apply in the European Union, the United Kingdom, or South Korea.

TTS and voice: Coqui TTS, Piper, and the forks that kept going

Coqui, the company behind Coqui TTS and XTTS, announced it was shutting down on January 3, 2024. The coqui-ai/TTS repo's last commit is from February 10, 2024, and the final release of the original TTS package on PyPI, 0.22.0 from December 2023, requires a Python version below 3.12. That ceiling alone rules it out on many current systems.

The project survived through a fork. The Idiap Research Institute maintains Coqui TTS as the coqui-tts package. Version 0.27.5 shipped in January 2026, it supports Python 3.10 through 3.14, commits were still landing on October 2, 2026, and it has added prebuilt macOS and Windows wheels, OpenVoice and kNN-VC voice conversion, and caching for cloned voices. Migration is a package swap, because the import path did not change:

pip uninstall -y TTS
pip install coqui-tts   # existing code keeps working: from TTS.api import TTS

The catch is licensing, not code. XTTS-v2 weights ship under the Coqui Public Model License, which permits non-commercial use, and the company that sold commercial licenses no longer exists. XTTS-v2 still logged about 6.7 million Hugging Face downloads in the past 30 days, but the license still rules it out for paid products.

Piper is the other notable move. The rhasspy/piper repo is archived with a one-line README: development has moved to OHF-Voice/piper1-gpl. Piper is alive there, with a commit on September 28, 2026, but the license changed from MIT to GPL-3.0, which matters if you embed it in closed-source software.

Bark (last commit April 2024), StyleTTS 2 (March 2024), Tortoise (November 2024), and MeloTTS (December 2024) have also gone quiet. The maintained replacements:

• Kokoro: 82M parameters, Apache 2.0, and around 11 million Hugging Face downloads in the past 30 days. Its last PyPI release, 0.9.4, dates from April 2025, which is fine for a small, finished model.
• Chatterbox: MIT licensed. Multilingual V3 (0.5B) clones voices across languages, Turbo (350M) targets low-latency English voice agents, and Nano (110M) runs 3x faster than realtime on 8 CPU cores, according to Resemble AI.
• F5-TTS: actively maintained with MIT code, but the pretrained weights are CC-BY-NC-4.0, the same commercial trap as XTTS.

More models are compared in our open-weight TTS roundup.

LLM inference: ExLlamaV2, TGI, and the quantization tools that folded

ExLlamaV2 shows why the GitHub archived badge is not enough. Since March 4, 2026, its README has opened with "This project is archived for now. Development continues on ExLlamaV3." The repository is not flagged as archived, so tooling that only checks the flag misses it. Its last release, v0.3.2, dates from July 2025.

The louder signal came from downstream. TabbyAPI, the recommended server for both ExLlama generations, removed its ExLlamaV2 backend on July 14, 2026, and its README now lists only EXL3 and FP16/BF16 models. ExLlamaV3 shipped v1.5.4 on October 3, 2026. EXL3 is a streamlined variant of the QTIP quantization method, and its converter takes an original Hugging Face model plus a target bitrate, so plan to re-quantize from source weights or download EXL3 quants:

git clone https://github.com/turboderp-org/exllamav3 && cd exllamav3
pip install exllamav3   # PyPI build needs the CUDA toolkit; prebuilt wheels are on the releases page
python convert.py -i ~/models/my-model -o ~/models/my-model-exl3 -w /tmp/exl3-work -b 4.0

One migration gotcha: ExLlamaV2 0.3.2 shipped ROCm wheels for AMD GPUs, while none of the 62 files attached to the ExLlamaV3 v1.5.4 release are ROCm builds. On AMD hardware, llama.cpp, with its HIP and Vulkan backends, is the practical landing spot.

Serving and quantization lost several more well-known names:

• Hugging Face Text Generation Inference is archived and in maintenance mode. Its README recommends vLLM, SGLang, llama.cpp, or MLX going forward.
• AutoAWQ is archived and described as officially deprecated; the vLLM project adopted its functionality into LLM Compressor.
• AutoGPTQ is archived, and its README points users to GPTQModel.
• Mistral's own reference library, mistral-inference, was archived in June 2026.

Existing AWQ and GPTQ checkpoints still load in vLLM, so already-quantized models are fine; only the tooling that creates new ones moved. The engines are compared in our inference engine overview.

3D generation: DreamGaussian, InstantMesh, and Zero-1-to-3

3D research code is usually released with a paper and then frozen, so judge these repos by whether they still install, not by commit frequency.

DreamGaussian (last commit January 2, 2024) optimizes 3D Gaussians per object and needs custom CUDA extensions, diff-gaussian-rasterization and simple-knn, compiled locally. Its README lists testing on torch 1.12 with CUDA 11.6 and torch 2.1 with CUDA 12.1; current PyTorch is 2.14.1. InstantMesh, TencentARC's feed-forward mesh generator, has had only README edits since May 2024 and pins transformers 4.34.1, diffusers 0.20.2, and gradio 3.41.2, against current releases of 5.19.0, 0.41.0, and 6.29.1. Zero-1-to-3, the ICCV 2023 novel-view synthesis model, has not changed since September 2023, and its demo needs about 22 GB of VRAM.

All three still work inside a dedicated environment built to their pins. None of them is where a new pipeline should start:

• TRELLIS.2 from Microsoft is MIT licensed, has 4B parameters, and outputs textured assets with PBR attributes (base color, roughness, metallic, opacity), including open surfaces and non-manifold geometry. It needs an NVIDIA GPU with at least 24 GB, and Linux is the only tested OS. On an H100, the README reports about 3 seconds per asset at 512^3 and about 17 seconds at 1024^3.
• Hunyuan3D 2.1 (June 2025) pairs a 3.3B shape model with a 2B PBR texture model and needs about 10 GB of VRAM for shape, 21 GB for texture, or 29 GB for both. It remains the newest general-purpose Hunyuan3D generator with downloadable weights; Tencent's later 2.5 and 3.x versions run only on its hosted platform and API. HunyuanVideo's territory exclusion applies here too.
• TripoSG is MIT licensed, has 1.5B parameters, and runs on 8 GB of VRAM. Its repo has been quiet since April 2025, proof that a finished model with a permissive license and modest requirements can be the safest pick.

For cluttered real-world photos, Meta's SAM 3D Objects (November 2025, SAM License) is worth a look. The full workflow is in our single-image-to-3D pipeline guide.

Detection, tracking, and depth: Grounding DINO, YOLO-World, Track Anything, MiDaS

Grounding DINO's official repo last changed on August 12, 2024, with a README note announcing Grounded SAM 2, and has 325 open issues and pull requests. IDEA Research's newer Grounding DINO 1.5 and DINO-X are offered through a cloud API, not as weights. Yet grounding-dino-base logged about 1.6 million Hugging Face downloads in the past 30 days, and Transformers supports it natively. Per the model card, text queries should be lowercase and end with a period:

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection

model_id = "IDEA-Research/grounding-dino-base"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id)

image = Image.open("street.jpg")
inputs = processor(images=image, text="a bicycle. a traffic cone.", return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
results = processor.post_process_grounded_object_detection(
    outputs, inputs.input_ids, threshold=0.4, text_threshold=0.3, target_sizes=[image.size[::-1]])

So stop cloning the original repo, keep using the model through Transformers if it meets your accuracy bar, and evaluate SAM 3 for new work. SAM 3 is an 848M-parameter model that detects, segments, and tracks every instance of a short text phrase in images and video. SAM 3.1 checkpoints with faster multi-object tracking arrived on March 27, 2026, and the repo had a commit on September 18, 2026. Checkpoints require a manual access request under the SAM License.

YOLO-World's repo last changed on February 26, 2025 and has 421 open issues and PRs. It remains usable through Ultralytics, which still ships YOLO-World weights but points open-vocabulary users toward YOLOE. YOLOE accepts text prompts, visual example prompts, or no prompt, using a built-in vocabulary of 4,585 names, and Ultralytics reports that YOLOE-v8s beats YOLO-Worldv2-S by 3.5 AP at a third of the training cost and 1.4x the inference speed:

from ultralytics import YOLOE

model = YOLOE("yoloe-26s-seg.pt")
model.set_classes(["forklift", "pallet"])
results = model.predict("warehouse.jpg")

Ultralytics is AGPL-3.0 and the original YOLO-World repo is GPL-3.0, so neither drops into a closed-source product; Ultralytics sells an Enterprise License for that case.

Track Anything paired the original Segment Anything with XMem for click-to-track video segmentation. Its last code change was in June 2023, and a December 2025 README edit points readers interested in human mesh generation to the authors' newer project, SAM-Body4D. SAM 2.1 does Track Anything's job natively with ungated Apache 2.0 weights, and SAM 3 adds text prompts. In fairness, the SAM 2 repo has been quiet since December 2024, but the model is stable and in Transformers.

MiDaS is formally archived, with its last commit on August 23, 2024. It is still on PyTorch Hub and still ships as the depth_midas preprocessor in controlnet_aux, so existing ControlNet workflows keep working. For new depth work:

• Depth Anything V2: the Small model is Apache 2.0; Base and Large are CC-BY-NC-4.0.
• Depth Anything 3 (November 2025) adds multi-view depth and camera pose; Small, Base, Metric, and Mono are Apache 2.0, Large and Giant CC BY-NC 4.0.
• Apple's Depth Pro estimates metric depth and focal length from one image, a 2.25-megapixel map in 0.3 seconds on a standard GPU, but its Hugging Face weights ship under Apple's research-only license.

Detection speed trade-offs are covered in our RF-DETR, YOLO, and SAM 3 comparison.

Vision language and code models: Qwen-VL and Code Llama

The original Qwen-VL repo has been idle since April 2024, and its checkpoints need trust_remote_code. That is succession more than abandonment. The line went Qwen2-VL, Qwen2.5-VL, then Qwen3-VL, all in one renamed repo where old Qwen2-VL links redirect, and in February 2026 vision moved into the main model line. Qwen3.5 is natively multimodal, and the Qwen team says it outperforms the Qwen3-VL models on reasoning, coding, agent, and visual understanding benchmarks. Its 0.8B to 9B sizes are Apache 2.0, and Qwen3.6 and Qwen3.8 followed.

Code Llama, Meta's Llama 2-based code family, lives in an archived repo, though the 7B weights still drew about 413,000 downloads in the past 30 days. For a local coding model today:

• Qwen3-Coder: the 30B-A3B instruct model is Apache 2.0 with a 256K context, and Qwen3-Coder-Next, built on Qwen3-Next-80B-A3B, targets coding agents and local development.
• Devstral Small 2 (December 2025): 24B parameters, Apache 2.0, a 256K context, and 68.0% on SWE-bench Verified according to Mistral.

How to spot an abandoned project before you depend on it

The projects above left a trail of warning signs. Check these before adding a dependency:

1. The archived flag, but not only the flag. MiDaS, Code Llama, Piper, TGI, AutoAWQ, and AutoGPTQ are flagged. ExLlamaV2 announces archival in its README with the flag off, and AnimateDiff, Grounding DINO, and the original Coqui TTS are not flagged at all.
2. The last commit to code. GitHub's updated date moves with stars, and pushed_at counts any branch. Track Anything had a README commit in December 2025 and no code change since June 2023.
3. Hard pins against current releases. diffusers==0.11.1 in a world on 0.41.0 means the authors stopped tracking upstream.
4. A Python ceiling. The original TTS package stops at Python 3.11.
5. A backlog without replies. Hundreds of open issues are normal for a popular repo; hundreds with no maintainer response for a year are not.
6. Downstream projects dropping support. TabbyAPI removing ExLlamaV2 said more than the ExLlamaV2 repo did.
7. The lab moved on. READMEs pointing to a new project or a paid API, as in Track Anything and Grounding DINO, show where the authors' attention went.
8. A busier fork. Sort forks by recent activity; the Idiap fork of Coqui TTS published its first release in April 2024, three months after the company closed.

Two commands cover the first two checks for any GitHub repo:

gh api repos/gaomingqi/Track-Anything --jq '{archived, pushed_at}'
gh api "repos/gaomingqi/Track-Anything/commits?per_page=5" \
  --jq '.[] | .commit.committer.date[0:10] + " " + (.commit.message | split("\n")[0])'

For a scored view, the OpenSSF Scorecard Maintained check gives full marks only for at least one commit per week over the previous 90 days, and the lowest score to archived repos:

export GITHUB_AUTH_TOKEN=<your token>
scorecard --repo=github.com/guoyww/AnimateDiff --checks Maintained --show-details

Calibrate for research code, though: nearly every paper repo fails that check by design. For those, ask whether the weights are permissively licensed and whether a maintained library such as Transformers, diffusers, Ultralytics, or a ComfyUI node pack has absorbed the model.

When an inactive project is still the right choice

Keep using a frozen project when:

• Its weights are permissively licensed and its output meets your bar, as with Kokoro, TripoSG, or Grounding DINO through Transformers.
• A maintained wrapper carries it: AnimateDiff in diffusers and ComfyUI, YOLO-World in Ultralytics, MiDaS in controlnet_aux.
• It runs offline in a pinned container you control.

Plan a migration when:

• Your serving stack has dropped it, as TabbyAPI did with ExLlamaV2.
• It blocks a Python, PyTorch, or CUDA upgrade, as the original Coqui TTS package and DreamGaussian do.
• It exposes a web demo to a network; the Gradio 3.x versions these repos pin miss years of fixes.
• The license blocks commercial use and nobody is left to sell an exception, as with XTTS-v2.

Common questions

Is AnimateDiff still worth using in 2026? For stylized loops on SD 1.5 checkpoints and LoRAs, yes, through diffusers or ComfyUI-AnimateDiff-Evolved rather than the original repo. For realistic motion or longer clips, Wan 2.2 or LTX-2.5 are built for the job.

Is Coqui TTS dead? The company closed in January 2024 and the original repo stopped a month later, but the Idiap fork (pip install coqui-tts) is maintained and supports Python 3.10 through 3.14. XTTS-v2 weights remain non-commercial.

What replaced ExLlamaV2? ExLlamaV3, served through TabbyAPI, which removed its ExLlamaV2 backend in July 2026. AMD users should look at llama.cpp, since ExLlamaV3 release builds target CUDA.

What is the best open source alternative to Grounding DINO? SAM 3 if you need masks and tracking from text prompts and accept its gated SAM License; YOLOE if you need real-time speed and AGPL-3.0 works for you. Grounding DINO itself still runs well through Transformers.

How do I tell if a GitHub AI project is abandoned? Check the archived flag, the last code commit on the default branch, dependency pins, maintainer replies, and whether downstream tools still support it. A README notice or a busier fork is often the first public sign.