The open-source AI models worth knowing in 2026
We run one small model ourselves right now — Qwen3-Coder-30B, on a single rented GPU. People assume that's the whole open-source model world. It isn't. There are hundreds of genuinely strong open-weight models released in the last year, across text, reasoning, vision, image, and voice — every one of them free to download and run yourself, no API key, no per-token bill. This post is the reading guide. For the exhaustive, constantly-refreshed list — 300+ models pulled live from the Hugging Face Hub API, sortable by downloads/license/category — see the live models table. This page explains what each category is and why it matters; that page is the actual data.
On cost, to be direct about it: everything below can be self-hosted for free — the only real expense is the GPU you rent to run it (or a laptop CPU for the small ones). Nowhere in this post am I suggesting we pay a per-token API bill to a model provider. Where a model is too large for our own hardware (GLM-5.2 is the example below), that's a "not on our infra, period" — not a "pay for it instead."
Reading the size column
Most of these are Mixture-of-Experts (MoE): a huge total parameter count, but only a fraction "active" per token — which is what actually decides your GPU memory and speed, not the headline number. Qwen3-Coder-30B is 30B total but only 3B active, which is why it runs at 100+ tokens/sec on one 48GB card. A 700B dense-equivalent model needs a small cluster. Always check active params, not just total.
General-purpose LLMsthe everyday chat-and-code workhorses
| Model | Size | License | Best for | Get it |
|---|---|---|---|---|
| Llama 3.1 / 3.3 | 8B–405B | Llama license (open weight, usage terms) | General-purpose, huge ecosystem/tooling support | meta-llama on HF |
| Qwen 2.5 / Qwen3 | 0.5B–235B (MoE) | Apache 2.0 | Best all-round open model line right now — chat, code, agentic tool-use | Qwen on HF |
| Qwen3-Coder-30B-A3B | 30B / 3B active | Apache 2.0 | Coding + tool-calling on a single 48GB GPU — what we run today | HF model card |
| DeepSeek-V3 | 671B / 37B active | MIT | Frontier-class quality at open weights, needs multi-GPU | deepseek-ai on HF |
| GLM-4.7-Flash | 31B | MIT | Same weight class as Qwen3-Coder — our next one to test | HF model card |
| GLM-5.2 | 753B | MIT | Frontier reasoning — needs ~10 GPUs, doesn't fit our one card. Off the table until infra changes, not something we'd pay per-token for | HF model card |
| Mistral Small / Large, Mixtral | 22B–123B | Apache 2.0 / Mistral license | Fast, strong in European languages, permissive | mistralai on HF |
| Gemma 2 / 3 | 2B–27B | Gemma license (open weight) | Small and cheap to run, good for edge/laptop use | google on HF |
| Phi-3 / Phi-4 | 3.8B–14B | MIT | Punches above its size — best small model for structured tasks | microsoft on HF |
| Kimi K2 | 1T / 32B active | Modified MIT | Very strong agentic/tool-use benchmark scores | moonshotai on HF |
Reasoning modelsslower, "thinks before answering" — for hard multi-step problems
| Model | Size | License | Notes | Get it |
|---|---|---|---|---|
| DeepSeek-R1 | 671B / 37B active | MIT | The model that proved open-source reasoning could match closed frontier models | HF model card |
| QwQ-32B | 32B | Apache 2.0 | Qwen's reasoning-tuned model, fits on one GPU | HF model card |
| Qwen3 (thinking mode) | up to 235B | Apache 2.0 | Same weights, toggle "thinking" on for harder prompts | Qwen on HF |
Vision-language models (VLMs)read an image or screenshot, answer in text
| Model | Size | License | Best for | Get it |
|---|---|---|---|---|
| Qwen2.5-VL | 3B–72B | Apache 2.0 | Strong OCR, document/UI understanding, video | HF model card |
| GLM-4.6V | ~9B–106B variants | MIT | Vision-language, pairs with the GLM text line | HF model card |
| InternVL 3 | 1B–78B | MIT / Apache 2.0 | Open research leader on many VLM benchmarks | OpenGVLab on HF |
| Llama 3.2 Vision | 11B / 90B | Llama license | Image understanding bolted onto the Llama line | meta-llama on HF |
| Pixtral 12B | 12B | Apache 2.0 | Mistral's compact vision model | HF model card |
Image generationtext-to-image, self-hostable
| Model | License | Notes | Get it |
|---|---|---|---|
| FLUX.1 (dev / schnell) | Non-commercial (dev) / Apache 2.0 (schnell) | Best open image quality right now; schnell is fast + fully free | black-forest-labs on HF |
| Stable Diffusion 3.5 / SDXL | Stability community license | Huge ecosystem of fine-tunes and LoRAs | stabilityai on HF |
| GLM-Image | MIT | Diffusers-based, ties into the same GLM family | HF model card |
Voice: TTS, cloning, and speech-to-textthe pipeline we already run parts of
| Model | License | Does | Get it |
|---|---|---|---|
| Whisper / faster-whisper | MIT | Speech-to-text, runs offline on CPU | GitHub |
| edge-tts | MIT (wrapper around a free Microsoft voice) | Natural TTS with zero setup — what Maya speaks with today | GitHub |
| Kyutai pocket-tts / Moshi | MIT / CC-BY | Voice cloning that runs on plain CPU, no GPU needed | GitHub |
| Coqui XTTS-v2 | CPML (non-commercial without a license) | High-quality multilingual voice cloning | HF model card |
| GLM-TTS / GLM-ASR-Nano | MIT | Zero-shot cloning + speech recognition, same GLM family | zai-org on HF |
Embeddings and searchthe layer underneath every RAG / knowledge-base system
| Model | License | Notes | Get it |
|---|---|---|---|
| BGE (BAAI) | MIT | Strong, small, cheap to run text embeddings | BAAI on HF |
| Qwen3-Embedding | Apache 2.0 | Pairs naturally with a Qwen text stack | HF model card |
| Nomic Embed | Apache 2.0 | Long-context embeddings, fully open training data too | nomic-ai on HF |
The engineering underneath — serving, not just weights
A model file alone does nothing. This is the part people skip, and it's the actual work:
- vLLM — the serving engine we run today; handles batching, quantization, and OpenAI-compatible endpoints so any app can call the model like it calls GPT.
- Ollama + llama.cpp — for models small enough to run fully local, no GPU rental at all.
- Hugging Face Hub — where every model above actually lives; also hosts free CPU/GPU Spaces for trying a model before you commit hardware to it.
- Quantization (
FP8,AWQ,GGUF) is what turns a 60GB model into a 17GB one that fits a smaller card — worth learning before assuming a model "doesn't fit."
Where this leaves us
We're running exactly one model (Qwen3-Coder-30B) on one GPU, and it isn't good enough for everyday use yet — that's a separate, real problem (no tool-wiring, no observability, not necessarily the wrong model choice). GLM-4.7-Flash is worth testing since it's the same weight class and free on hardware we already have, but it's one option among the 320+ on the live table, not a foregone conclusion. Everything past that — a second reasoning model, real vision, real voice cloning — is a GPU/cost decision, not a "does it exist" question. It exists. All of it, MIT or Apache-licensed, sitting on Hugging Face right now, for $0 to download.
References
Primary sources for everything above:
· Hugging Face model hub — search any name above directly
· Open LLM Leaderboard — independent benchmark comparisons
· vLLM · Ollama · llama.cpp
· The infra we run these on lives on my open-source stack post.