Open Source AI Models by Category: LLMs, Vision, Embeddings, Speech, and Image (2026)

TL;DR

Open source AI models are downloadable model weights covering language, vision, embedding, speech, and image tasks. Hugging Face hosts 2.96 million public model repositories. Most ship under Apache 2.0 or MIT, a few under custom terms with revenue thresholds, and almost none publish training data, which is why the accurate label is open weight.

Published September 1, 2026; updated September 7, 2026 with Hugging Face Hub usage data and an evaluation section on host fidelity and quantization. Every license, size, and context figure below was read from the model card or license file.

Open source AI models are models whose weights you can download and run. The label covers five categories that matter in production, and in each one the best open model now ships under MIT or Apache 2.0 with no strings, or under a custom license whose conditions bind only inference resellers. The open LLMs sit eight points under the closed frontier on the Artificial Analysis Intelligence Index v4.3 (45 vs 53), checked September 7, 2026. The open speech, embedding, and image models are the default choices in their categories, not the budget option.

Two sentences on the fine print, then the models. Almost none of these meet the OSI Open Source AI Definition, because none publish training data; the license text on the checkpoint is what governs what you can do. The full license table, the open-source vs open-weight breakdown, and the API vs self-host vs dedicated cost crossover live on the open source LLM guide.

How Many Open Source AI Models Are There, and Which Ones Get Used

The catalog is larger than any list, and usage is concentrated at the opposite end from the leaderboards. Hugging Face counted the whole Hub in State of Open Models: Summer 2026, published August 14, 2026.

Read the download split before you pick a model. The categories below list the strongest model in each modality, but 83% of downloads go to models under 1B parameters and 1% to models above 100B, mostly as quantized GGUF files pulled for local use. The permissive licenses are winning at the same time: of 178 Chinese releases above 20B parameters this year, 59% carry Apache 2.0 and 22% carry MIT, and Hugging Face notes almost none carry non-commercial restrictions. The custom licenses with revenue thresholds are a small, high-profile minority, and they bind companies reselling inference rather than teams shipping a product feature.

Coding and General LLMs

The largest category and the one with independent scoring. Artificial Analysis runs every model through one harness; the scores below are its Intelligence Index v4.3, checked September 7, 2026, with the closed leader (Claude Fable 5.1) at 53 for reference.

The four highlighted rows are served on Morph today. The table has one structural fact in it: every frontier-class open LLM is a mixture of experts with 13B to 104B active parameters, which is why 284B and 320B models serve at $0.3475 and $0.35 per million output tokens. Qwen 3.8 27B is the exception, a dense model that fits one GPU and scores one point under DeepSeek V4 Flash. Per-model pages: Kimi K3, GLM-5.3, GLM-5.3-Flash, DeepSeek V4 Flash, DeepSeek V4, gpt-oss. Ranked lists: best open source LLM and best open source coding model.

Vision-Language Models

A vision-language model takes images (and sometimes video) alongside text and answers in text. In 2026 the line between this category and the LLM category has mostly dissolved: three of the strongest open LLMs are natively multimodal, and the dedicated VLM families (Qwen3-VL) share an architecture with their text-only siblings.

For coding agents the useful VLM is the one that reads a screenshot of the rendered UI and edits the component. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series and is served on Morph with image input; Kimi K3's card lists video input as well. Qwen3-VL is the choice when you need Apache 2.0 and a vision model you can fine-tune. Gemma 4 31B is the single-GPU option. See MiniMax M3 for the video-input case.

Embedding and Reranking Models

Embedding models turn text into vectors for retrieval; rerankers score a query against candidate passages. This is the category where open models won outright: the top entries are small, permissively licensed, and run on CPU or a fraction of one GPU.

Pick by input length and dimension budget. Qwen3-Embedding-8B handles 32K tokens per input and lets you shrink the vector to 32 dimensions for cheap indexes. bge-m3 is the long-running multilingual default at 1024 dimensions and 8,192 tokens. nomic-embed v2 is the smallest and caps at 512 tokens, so it fits short-passage retrieval. All three include code among their languages. For local use, Ollama embedding models covers the quantized builds.

Speech: Transcription and Text to Speech

Two directions. Automatic speech recognition (audio to text) has two open families that set the reference points on the Hugging Face Open ASR Leaderboard. Text to speech now has open models under Apache 2.0 and MIT that are small enough to run per request.

Whisper large-v3 is the model to reach for when language coverage matters. Parakeet TDT 0.6B v3 is the model to reach for when throughput matters: a real-time factor above 3,000 means one GPU transcribes an hour of audio in about a second, and it handles files up to 24 minutes with full attention or 3 hours with local attention. On the output side, Kokoro is 82M parameters, about one twentieth the size of Whisper large-v3; Chatterbox Multilingual V3 is a 0.5B general-purpose model under MIT covering 23 languages.

Image Generation

The image category has the widest license spread, and the license, not the sample quality, is usually what decides. Two of the four leading models are Apache 2.0; the other two carry conditions that exclude most commercial products.

FLUX.1-schnell is the fastest path to a shippable image feature: Apache 2.0, 1 to 4 sampling steps, 12B parameters. Qwen-Image is the larger Apache 2.0 alternative at 20B parameters on an MMDiT architecture. FLUX.1-dev shares schnell's 12B architecture without the step distillation, but its license bars commercial use of the model itself. Stable Diffusion 3.5 Large is free only while your organization earns under $1M a year.

Evaluating One Before You Deploy It

A model card is a claim about a checkpoint. What you deploy is that checkpoint at some precision, served by some stack. Two measurements from 2026 show how much that changes the answer, and both are reproducible.

Precision: flat, then a cliff

Quesma benchmarked Qwen3.8 27B across five precisions on August 26, 2026. Q8_0 and 4-bit Q4_K_M show no measurable loss against the 55 GB BF16 reference, with Q4_K_M at 17 GB matching BF16's roughly 76% pass rate on Terminal-Bench 2.1. Two-bit drops to about 71% on the same agentic suite, and 1-bit builds fall to random chance on GPQA Diamond. Agentic and tool-use tasks degrade before single-turn knowledge does, so test the quantized build on the workload you actually run.

Host: the same weights behave differently

Moonshot publishes K2-Vendor-Verifier, which sends identical requests to every vendor serving its weights and compares tool-call behavior against the official API. On the K2 0905 run the official API returned schema-valid tool calls on all 1,274 successful calls, while the weakest deployments in the same table scored 76% and 73.13%. Ask a host which precision it serves and whether the checkpoint revision is pinned. The full breakdown, including OpenRouter's quality-first :exacto variant, is on the open source LLM guide.

Serving Them: One API First, Then Dedicated

Downloading a checkpoint is the easy part. The categories above split into two serving situations.

LLMs and VLMs: call them today

Morph serves four of the models above on one OpenAI-compatible endpoint at the full 1M context, with prefix caching on by default, a standby tier at 50% off, and an OpenAI-compatible Batch API at the same rate. GLM-5.3-Flash is the multimodal entry in the lineup and takes image input through the standard chat-completions message format.

from openai import OpenAI

client = OpenAI(base_url="https://api.morphllm.com/v1", api_key="YOUR_MORPH_API_KEY")

resp = client.chat.completions.create(
    model="morph-glm53flash",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "The header overlaps the nav on mobile. Fix the CSS."},
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
        ],
    }],
)
print(resp.choices[0].message.content)

Everything else: the open serving stacks, or dedicated capacity

Embedding, speech, and image models do not speak chat completions, and each has a mature open stack: Text Embeddings Inference or vLLM for embedding models, NVIDIA NeMo for Parakeet and faster-whisper for Whisper, diffusers for FLUX.1 and Qwen-Image. Run them on your own GPUs when volume is steady and the data is sensitive, or on an API when it is not. When you need reserved, isolated capacity behind a single endpoint, dedicated inference provisions B200 or B300 GPUs billed by the minute, with the calculator comparing the monthly cost against per-token spend. The self-serve catalog covers the LLM lineup; for a VLM, embedding, or speech model on dedicated hardware, talk to us.

FAQ

What are open source AI models?

Models whose trained weights are published for download, across language, vision, embedding, speech, and image tasks. Nearly all are open-weight rather than OSI open source, since the training data stays private; the license file on the checkpoint governs use.

Which open source AI model is best for each category?

LLMs: GLM-5.3 (Intelligence Index v4.3 score of 45) and Kimi K3 (44), GLM-5.3-Flash for cost (42, MIT), checked September 7, 2026. Vision: Qwen3-VL-235B (Apache 2.0) or GLM-5.3-Flash. Embeddings: Qwen3-Embedding-8B or bge-m3. Speech to text: Whisper large-v3 for coverage, Parakeet TDT 0.6B v3 for throughput. Text to speech: Kokoro or Chatterbox. Image: FLUX.1-schnell or Qwen-Image, both Apache 2.0.

Can I use them commercially?

Apache 2.0 and MIT models, yes, with no conditions. Kimi K3 and GLM-5.3 attach revenue thresholds that bind inference resellers. FLUX.1-dev is non-commercial. Stable Diffusion 3.5 is free only under $1M annual revenue. The license table has the clauses.

How many open source AI models are there?

Hugging Face counted 2.96 million public model repositories in August 2026, up from 2.43 million, plus 1 million datasets. Attention and usage diverge: models under 1B parameters take 83% of all-time downloads and models above 100B take 1%.

Why are Chinese AI models open source?

Distribution, monetized at the top of the funnel. Qwen's broad release strategy logged 2,045 million downloads in 2026 and 151,448 derivative models, against 37 million for Moonshot's frontier-only portfolio. Of 178 Chinese releases above 20B parameters this year, 59% carry Apache 2.0 and 22% carry MIT. The custom licenses on Kimi K3 and GLM-5.3 bind companies reselling inference above revenue thresholds, not teams shipping a feature.

How do open source AI models make money?

First-party inference on the lab's own API; license thresholds that trigger a paid agreement above a revenue or user line (Kimi K3 at $20M of model-as-a-service revenue, GLM-5.3 at $10B, Llama 4 at 700M monthly active users); tiered releases where the free tier is capped (Stable Diffusion 3.5 under $1M annual revenue, FLUX.1-dev non-commercial while FLUX.1-schnell is Apache 2.0); and the enterprise pull that comes from being the default checkpoint.

How do I evaluate open source AI models?

Evaluate the deployed artifact. A 4-bit build of Qwen3.8 27B matches BF16 on Terminal-Bench 2.1 while 2-bit loses about five points and 1-bit collapses, and the same weights served by different hosts score between 73% and 100% on tool-call schema validity. Fix the precision, pin the revision, then run your own task suite.

Where do I download them?

Hugging Face, under each developer's organization; the Sources list below links every model card. For quantized local builds of the smaller LLMs, see best Ollama models.

How do I serve them in production?

LLMs and VLMs: an OpenAI-compatible API (Morph serves four at 1M context) or vLLM and SGLang on your own GPUs. Embedding, speech, and image models: their own open serving stacks, or dedicated inference when you need reserved capacity.

Related Resources

• Open Source LLM: Licenses, Models, and One API to Run Them
• Best Open Source LLMs (2026), Ranked
• Best Ollama Models and Ollama Embedding Models
• LLM Inference: Cost, Speed, Caching, and GPU Capacity
• Dedicated Inference and the cost calculator
• Morph open models