{"ok":true,"source":"tensorfeed.ai","lastUpdated":"2026-09-14","count":27,"tools":[{"id":"ollama","name":"Ollama","vendor":"Ollama","category":"runtime","language":"Go","license":"MIT","starsK":180.9,"version":"0.34.x","released":"2023-07","features":["model library","one-line install","OpenAI-compatible API","GGUF native"],"url":"https://ollama.com","github":"https://github.com/ollama/ollama","notes":"Most-installed local LLM runtime. Mac / Linux / Windows. Curated model library; one command to pull and run. Past 180,000 GitHub stars as of September 2026."},{"id":"lm-studio","name":"LM Studio (Bionic)","vendor":"LM Studio","category":"runtime","language":"TypeScript / C++","license":"Proprietary (free tier)","starsK":0,"version":"Bionic 1.1.x","released":"2023-05","features":["GUI","GGUF + MLX","OpenAI + Anthropic-compatible local server","agent with skills and subagents","LM Link multi-device"],"url":"https://lmstudio.ai","github":"https://github.com/lmstudio-ai","notes":"Polished desktop app for running local models, relaunched on July 16, 2026 as LM Studio Bionic, an agent for open models. The free tier still runs local LLMs through llama.cpp and MLX; Bionic+ ($20/mo) and Pro ($100/mo) add US-hosted open models with zero data retention. App is closed source; SDKs and the lms CLI are open."},{"id":"llama-cpp","name":"llama.cpp","vendor":"ggml-org (Georgi Gerganov)","category":"runtime","language":"C++","license":"MIT","starsK":128.2,"version":"0.4.x (plus rolling bNNNNN builds)","released":"2023-03","features":["GGUF","CPU + GPU + Metal","quantization","server mode"],"url":"https://github.com/ggml-org/llama.cpp","github":"https://github.com/ggml-org/llama.cpp","notes":"The reference open-source LLM inference engine. Originated the GGUF format and powers Ollama, LM Studio, Jan, and most other local runtimes. Now lives under the ggml-org GitHub org and publishes semver releases (0.4.0 on September 4, 2026) alongside the rolling build tags."},{"id":"vllm","name":"vLLM","vendor":"UC Berkeley + community","category":"inference-server","language":"Python / CUDA","license":"Apache-2.0","starsK":91.7,"version":"0.29.x","released":"2023-06","features":["PagedAttention","continuous batching","tensor parallelism","OpenAI-compatible API","speculative decoding"],"url":"https://github.com/vllm-project/vllm","github":"https://github.com/vllm-project/vllm","notes":"Highest-throughput open-source inference server. Production default for self-hosted multi-GPU LLM serving, and one of the engines Hugging Face now recommends in place of its own TGI."},{"id":"sglang","name":"SGLang","vendor":"LMSYS Org + community","category":"inference-server","language":"Python / CUDA","license":"Apache-2.0","starsK":35.9,"version":"0.5.x","released":"2024-01","features":["RadixAttention","structured generation","MoE-optimized","tensor + expert parallelism"],"url":"https://github.com/sgl-project/sglang","github":"https://github.com/sgl-project/sglang","notes":"High-performance inference engine that often matches or beats vLLM throughput. Strong on large MoE models."},{"id":"tgi","name":"Text Generation Inference (TGI)","vendor":"Hugging Face","category":"inference-server","language":"Rust","license":"Apache-2.0","starsK":10.9,"version":"3.3.x (archived)","released":"2023-04","features":["continuous batching","tensor parallelism","speculative decoding","LoRA serving"],"url":"https://github.com/huggingface/text-generation-inference","github":"https://github.com/huggingface/text-generation-inference","notes":"Hugging Face's former production inference server. Now in maintenance mode and archived on GitHub (last release 3.3.7, December 2025); Hugging Face recommends vLLM, SGLang, llama.cpp, or MLX going forward. Listed for existing deployments only."},{"id":"tei","name":"Text Embeddings Inference","vendor":"Hugging Face","category":"inference-server","language":"Rust","license":"Apache-2.0","starsK":5,"version":"1.9.x","released":"2024-01","features":["embedding model serving","reranker support","GPU + CPU","OpenAI-compatible API"],"url":"https://github.com/huggingface/text-embeddings-inference","github":"https://github.com/huggingface/text-embeddings-inference","notes":"Production embedding server. Drop-in for the OpenAI embeddings API. Faster than vanilla sentence-transformers. Still maintained, unlike its sibling TGI."},{"id":"mlx","name":"MLX","vendor":"Apple","category":"runtime","language":"Python / C++","license":"MIT","starsK":28.4,"version":"0.32.x","released":"2023-12","features":["Apple Silicon native","unified memory","lazy execution","distributed training"],"url":"https://github.com/ml-explore/mlx","github":"https://github.com/ml-explore/mlx","notes":"Apple's ML framework, optimized for M-series chips. Strong inference and fine-tuning on high-memory Macs, and the second engine inside LM Studio."},{"id":"mlc-llm","name":"MLC LLM","vendor":"MLC AI Team","category":"runtime","language":"Python / C++","license":"Apache-2.0","starsK":23.2,"version":"0.20.x","released":"2023-04","features":["cross-platform compilation","WebGPU","Android + iOS deployment","TVM-based"],"url":"https://github.com/mlc-ai/mlc-llm","github":"https://github.com/mlc-ai/mlc-llm","notes":"Compile LLMs to run anywhere: web browsers (WebGPU), iOS, Android, embedded. Best fit for on-device inference research."},{"id":"exllamav3","name":"ExLlamaV3","vendor":"turboderp","category":"runtime","language":"Python / C++","license":"MIT","starsK":1.4,"version":"1.5.x","released":"2025-04","features":["EXL3 quantization","tensor + expert parallelism","CPU offloading for MoE","speculative decoding"],"url":"https://github.com/turboderp-org/exllamav3","github":"https://github.com/turboderp-org/exllamav3","notes":"Successor to ExLlamaV2, which is now archived. Runs quantized local models on consumer GPUs with the QTIP-based EXL3 format, flexible tensor and expert parallelism, and CPU offloading for large MoE models. TabbyAPI is the recommended OpenAI-compatible server."},{"id":"kobold-cpp","name":"KoboldCpp","vendor":"LostRuins + community","category":"runtime","language":"C++","license":"AGPL-3.0","starsK":11.7,"version":"1.120","released":"2023-04","features":["GGUF","creative writing UI","image gen","multimodal"],"url":"https://github.com/LostRuins/koboldcpp","github":"https://github.com/LostRuins/koboldcpp","notes":"Single-binary local LLM with creative-writing UI. Popular in roleplay / fiction communities."},{"id":"unsloth","name":"Unsloth","vendor":"Unsloth AI","category":"fine-tuning","language":"Python","license":"Apache-2.0","starsK":76.1,"version":"2026.9","released":"2023-11","features":["2x faster training","up to 70% less VRAM","LLM, diffusion, TTS, and embedding models","free Colab notebooks"],"url":"https://unsloth.ai","github":"https://github.com/unslothai/unsloth","notes":"Fast open fine-tuning library that now covers LLMs, diffusion, TTS, and embedding models, with free Colab notebooks and an Unsloth Studio UI. Past 76,000 GitHub stars as of September 2026."},{"id":"llamafactory","name":"LlamaFactory","vendor":"hiyouga + community","category":"fine-tuning","language":"Python","license":"Apache-2.0","starsK":74.8,"version":"0.9.x","released":"2023-05","features":["100+ LLMs and VLMs","zero-code CLI","Gradio web UI","full, freeze, LoRA, and QLoRA tuning"],"url":"https://llamafactory.readthedocs.io","github":"https://github.com/hiyouga/LlamaFactory","notes":"Unified fine-tuning toolkit for 100+ LLMs and VLMs with a zero-code CLI and web UI (ACL 2024 paper). Cited by Amazon, NVIDIA, and Alibaba Cloud in their own tutorials."},{"id":"trl","name":"TRL","vendor":"Hugging Face","category":"fine-tuning","language":"Python","license":"Apache-2.0","starsK":19.3,"version":"1.13.x","released":"2020-03","features":["SFT","GRPO","DPO","KTO","Transformers-native"],"url":"https://hf.co/docs/trl","github":"https://github.com/huggingface/trl","notes":"Hugging Face's post-training library for transformer language models, with trainers for supervised fine-tuning, GRPO, DPO, and KTO built on the Transformers ecosystem. Now past 1.0."},{"id":"axolotl","name":"Axolotl","vendor":"OpenAccess AI Collective","category":"fine-tuning","language":"Python","license":"Apache-2.0","starsK":12.5,"version":"0.19.x","released":"2023-05","features":["YAML-driven config","LoRA + QLoRA + full SFT + DPO","multi-GPU","large-scale recipes"],"url":"https://axolotl.ai","github":"https://github.com/axolotl-ai-cloud/axolotl","notes":"YAML-config fine-tuning toolkit. A common default for production-scale custom training runs."},{"id":"torchtune","name":"TorchTune","vendor":"Meta","category":"fine-tuning","language":"Python","license":"BSD-3-Clause","starsK":5.8,"version":"0.6.x (no longer maintained)","released":"2024-04","features":["PyTorch-native","distributed training","LoRA + QLoRA + full","composable recipes"],"url":"https://github.com/meta-pytorch/torchtune","github":"https://github.com/meta-pytorch/torchtune","notes":"Meta's PyTorch-native fine-tuning library. No longer actively maintained: development wound down in 2025 and the last release was 0.6.1 in April 2025. The repo now lives under meta-pytorch. Listed for existing users only."},{"id":"open-webui","name":"Open WebUI","vendor":"Open WebUI Inc.","category":"ui","language":"Python / Svelte","license":"Open WebUI License (BSD-3-Clause plus branding clause)","starsK":151.9,"version":"0.11.x","released":"2023-10","features":["ChatGPT-style UI","Ollama integration","RAG","function calling","multi-user"],"url":"https://openwebui.com","github":"https://github.com/open-webui/open-webui","notes":"Most-deployed local LLM UI. Self-hostable; supports Ollama, OpenAI, and any compatible API. Multi-user out of the box. Its license bars removing Open WebUI branding in deployments above 50 users without permission."},{"id":"librechat","name":"LibreChat","vendor":"Danny Avila","category":"ui","language":"TypeScript / Node","license":"MIT","starsK":43.4,"version":"0.8.x","released":"2023-04","features":["multi-provider","plugins","MCP support","workflows","agents"],"url":"https://www.librechat.ai","github":"https://github.com/danny-avila/LibreChat","notes":"Self-hosted multi-LLM chat UI with strong agent and MCP integration. Production-ready alternative to ChatGPT for orgs that want control."},{"id":"jan","name":"Jan","vendor":"Menlo Research","category":"ui","language":"TypeScript / Tauri","license":"Apache-2.0","starsK":44.5,"version":"0.8.x","released":"2024-01","features":["offline-first","bundled llama.cpp engine","MCP integration","OpenAI-compatible local API"],"url":"https://jan.ai","github":"https://github.com/janhq/jan","notes":"Open-source ChatGPT replacement that downloads and runs LLMs locally with full control and privacy. Bundles a llama.cpp engine, exposes an OpenAI-compatible server on localhost:1337, and supports MCP. Licensed Apache-2.0; the repo now lives at janhq/jan."},{"id":"lm-eval-harness","name":"lm-evaluation-harness","vendor":"EleutherAI","category":"eval","language":"Python","license":"MIT","starsK":14,"version":"0.4.x","released":"2020-11","features":["200+ benchmark tasks","multi-backend (HF, vLLM, OpenAI)","reproducible scoring","few-shot evaluation"],"url":"https://github.com/EleutherAI/lm-evaluation-harness","github":"https://github.com/EleutherAI/lm-evaluation-harness","notes":"The reference open-source eval harness and the standard for \"can I reproduce a benchmark score.\""},{"id":"inspect-ai","name":"Inspect AI","vendor":"UK AI Security Institute","category":"eval","language":"Python","license":"MIT","starsK":2.8,"version":"0.3.x","released":"2024-05","features":["agentic eval","tool-use scoring","human-in-the-loop","reproducible runs"],"url":"https://inspect.aisi.org.uk","github":"https://github.com/UKGovernmentBEIS/inspect_ai","notes":"The UK AI Security Institute's eval framework, designed for agentic and tool-use evaluations. Ships frequent 0.3.x point releases (0.3.263 on September 4, 2026)."},{"id":"opik","name":"Opik","vendor":"Comet","category":"observability","language":"Python / TypeScript","license":"Apache-2.0","starsK":22,"version":"2.2.x","released":"2024-09","features":["LLM tracing","eval scoring","feedback loops","self-host or cloud"],"url":"https://www.comet.com/site/products/opik/","github":"https://github.com/comet-ml/opik","notes":"OSS LLM observability and evaluation. Trace any LangChain / LlamaIndex / OpenAI run. Free self-host; paid cloud tier. Grew past 20,000 GitHub stars on its 2.x line."},{"id":"langfuse","name":"Langfuse","vendor":"Langfuse (ClickHouse)","category":"observability","language":"TypeScript / Python","license":"MIT (ee/ directories excluded)","starsK":34.6,"version":"4.x","released":"2023-08","features":["LLM tracing","prompt management","datasets + evals","self-host"],"url":"https://langfuse.com","github":"https://github.com/langfuse/langfuse","notes":"Most-deployed open LLM observability platform, now part of ClickHouse. Langfuse v4 made tracing real-time, which the team says is up to 165x faster. Free self-host; managed cloud tier. Strong fit for LangGraph / LangChain stacks."},{"id":"comfyui","name":"ComfyUI","vendor":"Comfy Org","category":"ui","language":"Python","license":"GPL-3.0","starsK":133,"version":"0.35.x","released":"2023-01","features":["node-graph UI","diffusion model support","video","3D extensions","custom nodes"],"url":"https://www.comfy.org","github":"https://github.com/Comfy-Org/ComfyUI","notes":"Node-based engine and UI for diffusion models. The default for advanced image and video workflow construction. Repo moved to the Comfy-Org GitHub org."},{"id":"sd-webui","name":"AUTOMATIC1111 / Stable Diffusion WebUI","vendor":"AUTOMATIC1111","category":"ui","language":"Python","license":"AGPL-3.0","starsK":164.9,"version":"1.10.x","released":"2022-08","features":["Stable Diffusion","extensions","LoRA","inpainting","animation"],"url":"https://github.com/AUTOMATIC1111/stable-diffusion-webui","github":"https://github.com/AUTOMATIC1111/stable-diffusion-webui","notes":"Most-starred Stable Diffusion UI with a massive plugin ecosystem, but development has stalled: the last release was 1.10.1 in February 2025. New workflows have largely moved to ComfyUI."},{"id":"onnxruntime","name":"ONNX Runtime","vendor":"Microsoft","category":"edge","language":"C++ / Python / Java","license":"MIT","starsK":21.8,"version":"1.30.x","released":"2018-12","features":["cross-platform","mobile + web","quantization","CPU + GPU"],"url":"https://onnxruntime.ai","github":"https://github.com/microsoft/onnxruntime","notes":"Production ONNX runtime. Best fit for deploying small models to mobile, browser, and embedded targets."},{"id":"tinygrad","name":"tinygrad","vendor":"tinycorp / George Hotz","category":"training","language":"Python","license":"MIT","starsK":33.6,"version":"0.14.x","released":"2020-11","features":["minimal ML framework","multiple accelerators","lazy compute graph"],"url":"https://tinygrad.org","github":"https://github.com/tinygrad/tinygrad","notes":"Minimalist deep learning framework. Backs Tinybox products (consumer-grade compute boxes)."}]}