2026 · 07 · 23 Models

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."

"Open-source" doesn't mean "runs on a laptop." A model's parameter count decides its GPU bill. That's the first filter, before you even ask if it's good.

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 LLMs
the everyday chat-and-code workhorses

ModelSizeLicenseBest forGet it
Llama 3.1 / 3.38B–405BLlama license (open weight, usage terms)General-purpose, huge ecosystem/tooling supportmeta-llama on HF
Qwen 2.5 / Qwen30.5B–235B (MoE)Apache 2.0Best all-round open model line right now — chat, code, agentic tool-useQwen on HF
Qwen3-Coder-30B-A3B30B / 3B activeApache 2.0Coding + tool-calling on a single 48GB GPU — what we run todayHF model card
DeepSeek-V3671B / 37B activeMITFrontier-class quality at open weights, needs multi-GPUdeepseek-ai on HF
GLM-4.7-Flash31BMITSame weight class as Qwen3-Coder — our next one to testHF model card
GLM-5.2753BMITFrontier reasoning — needs ~10 GPUs, doesn't fit our one card. Off the table until infra changes, not something we'd pay per-token forHF model card
Mistral Small / Large, Mixtral22B–123BApache 2.0 / Mistral licenseFast, strong in European languages, permissivemistralai on HF
Gemma 2 / 32B–27BGemma license (open weight)Small and cheap to run, good for edge/laptop usegoogle on HF
Phi-3 / Phi-43.8B–14BMITPunches above its size — best small model for structured tasksmicrosoft on HF
Kimi K21T / 32B activeModified MITVery strong agentic/tool-use benchmark scoresmoonshotai on HF

Reasoning models
slower, "thinks before answering" — for hard multi-step problems

ModelSizeLicenseNotesGet it
DeepSeek-R1671B / 37B activeMITThe model that proved open-source reasoning could match closed frontier modelsHF model card
QwQ-32B32BApache 2.0Qwen's reasoning-tuned model, fits on one GPUHF model card
Qwen3 (thinking mode)up to 235BApache 2.0Same weights, toggle "thinking" on for harder promptsQwen on HF

Vision-language models (VLMs)
read an image or screenshot, answer in text

ModelSizeLicenseBest forGet it
Qwen2.5-VL3B–72BApache 2.0Strong OCR, document/UI understanding, videoHF model card
GLM-4.6V~9B–106B variantsMITVision-language, pairs with the GLM text lineHF model card
InternVL 31B–78BMIT / Apache 2.0Open research leader on many VLM benchmarksOpenGVLab on HF
Llama 3.2 Vision11B / 90BLlama licenseImage understanding bolted onto the Llama linemeta-llama on HF
Pixtral 12B12BApache 2.0Mistral's compact vision modelHF model card

Image generation
text-to-image, self-hostable

ModelLicenseNotesGet it
FLUX.1 (dev / schnell)Non-commercial (dev) / Apache 2.0 (schnell)Best open image quality right now; schnell is fast + fully freeblack-forest-labs on HF
Stable Diffusion 3.5 / SDXLStability community licenseHuge ecosystem of fine-tunes and LoRAsstabilityai on HF
GLM-ImageMITDiffusers-based, ties into the same GLM familyHF model card

Voice: TTS, cloning, and speech-to-text
the pipeline we already run parts of

ModelLicenseDoesGet it
Whisper / faster-whisperMITSpeech-to-text, runs offline on CPUGitHub
edge-ttsMIT (wrapper around a free Microsoft voice)Natural TTS with zero setup — what Maya speaks with todayGitHub
Kyutai pocket-tts / MoshiMIT / CC-BYVoice cloning that runs on plain CPU, no GPU neededGitHub
Coqui XTTS-v2CPML (non-commercial without a license)High-quality multilingual voice cloningHF model card
GLM-TTS / GLM-ASR-NanoMITZero-shot cloning + speech recognition, same GLM familyzai-org on HF

Embeddings and search
the layer underneath every RAG / knowledge-base system

ModelLicenseNotesGet it
BGE (BAAI)MITStrong, small, cheap to run text embeddingsBAAI on HF
Qwen3-EmbeddingApache 2.0Pairs naturally with a Qwen text stackHF model card
Nomic EmbedApache 2.0Long-context embeddings, fully open training data toonomic-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:

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.

Trying to pick a model without a GPU budget?

I map open-weight models onto real GPU cost and real business problems for a living. If you're stuck choosing between "which model" and "can I afford to run it," tell me what you're building.

Book a call →
← All notes