Every model here can be downloaded and run on your own hardware. That is where the similarity ends, because open weights and a permissive licence are different things and the difference is what decides whether you can ship. MIT at 320 billion parameters and a bespoke vendor licence at 753 billion are not the same offer, and the table says which is which.
Selected as every downloadable model in our index, with licences taken from the model card and the Hugging Face API rather than from a launch post. Where a card does not name a standard licence we write what it says rather than guessing at an equivalent. Download and like counts are from the Hugging Face API on 8 September 2026 and are a proxy for adoption, not for quality.
| Model | Vendor | Licence | Params | Context | Likes | Downloads / mo |
|---|---|---|---|---|---|---|
| Open-weight frontier10 | ||||||
Qwen3.8-27B Qwen/Qwen3.8-27B Beats Claude Opus 4.6 Max on SWE-bench Pro and OSWorld in Alibaba's table, at 27B under Apache 2.0. Second most-liked model on Hugging Face. | Alibaba | Apache 2.0 | 27.8B27.8B dense active | 262K, 1M max | 14.2k | 6.4M |
Kimi K3 moonshotai/Kimi-K3 The largest open-weight model shipped, at 2.8 trillion parameters, and third on the Artificial Analysis index by third-party trackers. | Moonshot | Modified MIT | 2.8T | 1M | 11.2k | 2.4M |
DeepSeek-V4-Pro deepseek-ai/DeepSeek-V4-Pro The DeepSeek flagship, MIT licensed. Second on CyberGym behind GLM-5.3, and a long-standing favourite for self-hosted agent work. | DeepSeek | MIT | — | — | 5.5k | 625.9k |
gpt-oss-120b openai/gpt-oss-120b OpenAI's open-weight release under Apache 2.0. Notable less for its scores than for existing at all, and 5.2M downloads a month. | OpenAI | Apache 2.0 | 120B | — | 5.2k | 5.2M |
Qwen3.8-Flash-Next Qwen/Qwen3.8-Flash-Next Beats Claude Opus by 22 points on AndroidWorld device automation. Frontier scores at roughly 6B inference cost, on a 180B machine. | Alibaba | qwen-community-1.0 | 180B6B active | 262K, 1M max | 5.0k | 474.7k |
DeepSeek-V4-Flash-0731 deepseek-ai/DeepSeek-V4-Flash-0731 MIT at 304B and 4.5M downloads a month. Widely described as the strongest cost-and-agent option among downloadable weights. | DeepSeek | MIT | 304B | — | 3.9k | 4.5M |
gemma-4-31B-it google/gemma-4-31B-it Google's open-weight line, 8.6M downloads a month. The gemma-4 family dominates the any-to-any category on Hugging Face. | Gemma Terms of Use | 31B | — | 3.7k | 8.6M | |
GLM-5.3-Flash zai-org/GLM-5.3-Flash MIT at 320B, multimodal, and four points behind its own flagship. The most permissive licence attached to a model this size. | Z.ai | MIT | 320B18B active | 300K | 2.1k | 784.0k |
GLM-5.3 zai-org/GLM-5.3 Level with GPT-5.6 Sol on Terminal-Bench 2.1 at a seventh of the input price, and state of the art on CyberGym vulnerability discovery. | Z.ai | glm-5.3, bespoke | 753B8 experts / token active | 1M | 1.7k | 442.1k |
Muse Glimmer 30B Meta's move to Apache 2.0 and to a dense architecture, away from the Llama Community Licence and mixture-of-experts. | Meta | Apache 2.0 | 30B | — | — | — |
| Small and on-device4 | ||||||
Qwen3-0.6B Qwen/Qwen3-0.6B 21M downloads a month, more than any frontier open model. Small enough for on-device, browser and edge-runtime inference. | Alibaba | Apache 2.0 | 0.6B | — | 1.6k | 21.3M |
Qwen3-8B Qwen/Qwen3-8B The workhorse size. Still 13M downloads a month a year after release, and where most self-hosted pilots start. | Alibaba | Apache 2.0 | 8B | — | 1.4k | 13.0M |
Spark-X2.5-4B XHToken/Spark-X2.5-4B Top of the Hugging Face trending list on the day we checked, on 697 likes against 7.2k downloads. Attention arriving well ahead of adoption. | XHToken | See model card | 4B | — | 699 | 7.2k |
MiniCPM5-2B openbmb/MiniCPM5-2B Trending on 159 likes with 13 downloads, which is what a release looks like in its first hours. Worth watching rather than deploying. | OpenBMB | See model card | 2B | — | 167 | 13 |
| Vision-language and OCR4 | ||||||
Unlimited-OCR baidu/Unlimited-OCR 2.7M downloads and 4.2k likes. Document extraction as a dedicated model rather than a prompt to a general vision-language model. | Baidu | See model card | — | — | 4.2k | 2.7M |
Qwen3-VL-8B-Instruct Qwen/Qwen3-VL-8B-Instruct 14.5M downloads a month. Fits on a single accelerator, which makes it the open-weight default for document and screen work. | Alibaba | Apache 2.0 | 8B | — | 1.1k | 14.5M |
DeepSeek-V4-Flash-Vision-Exp deepseek-ai/DeepSeek-V4-Flash-Vision-Exp DeepSeek's experimental vision checkpoint at 305B, trending within days of release. Experimental is in the name, so treat it that way. | DeepSeek | MIT | 305B | — | 789 | 251.6k |
PaddleOCR VL 1.5 Tops OmniDocBench at 94.37 overall, ahead of every general model on that test. A parser rather than a reasoner. | Open weights | Apache 2.0 | — | — | — | — |
| Image generation4 | ||||||
FLUX.1-dev black-forest-labs/FLUX.1-dev The most-liked model on Hugging Face, full stop, at 14,508 likes. Note the licence: the dev weights are non-commercial. | Black Forest Labs | Non-commercial, dev | — | — | 14.5k | 774.3k |
FLUX.1-schnell black-forest-labs/FLUX.1-schnell The commercially usable FLUX. Apache 2.0 where the dev weights are not, which is the version most products actually ship. | Black Forest Labs | Apache 2.0 | — | — | 5.7k | 717.8k |
Z-Image-Turbo Tongyi-MAI/Z-Image-Turbo 5.2k likes and 690k downloads. The Alibaba-adjacent entry in open image generation, and a genuine FLUX alternative. | Tongyi-MAI | See model card | — | — | 5.2k | 689.5k |
Krea-2-Turbo krea/Krea-2-Turbo Top of the Hugging Face trending list for text-to-image on the day we checked, ahead of FLUX on momentum if not on total likes. | Krea | See model card | — | — | 1.1k | 76.2k |
| Video generation3 | ||||||
MiniMax-H3 MiniMaxAI/MiniMax-H3 5M downloads and 5k likes, with a whole ecosystem of turbo, quantised and LoRA derivatives already on Hugging Face. | MiniMaxAI | See model card | 33B | — | 5.0k | 5.0M |
LTX-2.5 Lightricks/LTX-2.5 1.6M downloads and 3.1k likes. The other serious open video model, and the one with the longer track record. | Lightricks | See model card | — | — | 3.1k | 1.6M |
Wan2.2-I2V-A14B Wan-AI/Wan2.2-I2V-A14B Image to video at 14B under Apache 2.0, with a large ComfyUI following. The permissively licensed option in open video. | Alibaba | Apache 2.0 | 14B | — | 794 | 10.2k |
| Speech to text6 | ||||||
whisper-large-v3 openai/whisper-large-v3 The model that made transcription a commodity. Still 6.2k likes and 5M downloads, and still the answer when audio cannot leave your network. | OpenAI | Apache 2.0 | — | — | 6.2k | 5.0M |
speaker-diarization-3.1 pyannote/speaker-diarization-3.1 9.1M downloads a month for the job the transcription models do not do: working out who was speaking. | pyannote | MIT | — | — | 3.4k | 9.1M |
whisper-large-v3-turbo openai/whisper-large-v3-turbo 6.9M downloads a month, more than the model it distils. The default self-hosted transcription choice on constrained hardware. | OpenAI | MIT | — | — | 3.3k | 6.9M |
parakeet-tdt-0.6b-v3 nvidia/parakeet-tdt-0.6b-v3 Very fast transcription at 0.6B. The pick when throughput per GPU matters more than the last point of accuracy. | NVIDIA | CC-BY-4.0 | 0.6B | — | 1.1k | 745.8k |
Qwen3-ASR-1.7B Qwen/Qwen3-ASR-1.7B 3.3M downloads a month. Open-weight transcription at a size you can genuinely self-host on modest hardware. | Alibaba | Apache 2.0 | 1.7B | — | 1.1k | 3.3M |
Voxtral-Mini-4B-Realtime mistralai/Voxtral-Mini-4B-Realtime-2602 Mistral's realtime speech model under Apache 2.0, at 2.2M downloads. The European option where that matters for procurement. | Mistral | Apache 2.0 | 4B | — | 973 | 2.2M |
| Text to speech3 | ||||||
Kokoro-82M hexgrad/Kokoro-82M 6.8k likes and 11.5M downloads at 82 million parameters. Proof that speech synthesis does not need a large model. | hexgrad | Apache 2.0 | 82M | — | 6.8k | 11.5M |
Qwen3-TTS-CustomVoice Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice 2.6M downloads. Voice cloning on open weights, which is a capability and a governance question in equal measure. | Alibaba | Apache 2.0 | 1.7B | — | 1.9k | 2.6M |
VoxCPM2 openbmb/VoxCPM2 1.6k likes on 399k downloads, and rising fast. One of the stronger recent open synthesis releases. | OpenBMB | See model card | — | — | 1.6k | 399.4k |
| Embeddings4 | ||||||
all-MiniLM-L6-v2 sentence-transformers/all-MiniLM-L6-v2 251 million downloads a month, the most-downloaded model on Hugging Face by a factor of three. Still the default first embedder. | sentence-transformers | Apache 2.0 | 22.7M | — | 5.6k | 251.4M |
BAAI/bge-m3 BAAI/bge-m3 37.7M downloads. Multilingual, multi-granularity, and the usual upgrade from MiniLM once retrieval quality starts to matter. | BAAI | MIT | — | — | 3.5k | 37.7M |
embeddinggemma-300m google/embeddinggemma-300m Google's small embedder, trending hard at 2.3M downloads. Built for on-device retrieval where the vectors never leave the phone. | Gemma Terms of Use | 300M | — | 1.9k | 2.3M | |
nomic-embed-text-v1.5 nomic-ai/nomic-embed-text-v1.5 16.2M downloads, with Matryoshka dimensions so you can trade vector size against recall without re-embedding. | Nomic | Apache 2.0 | — | — | 902 | 16.2M |
| Reranking3 | ||||||
bge-reranker-v2-m3 BAAI/bge-reranker-v2-m3 18M downloads. Reranking is the cheapest quality win in a RAG pipeline and this is the model most teams reach for first. | BAAI | Apache 2.0 | — | — | 1.2k | 18.0M |
ms-marco-MiniLM-L6-v2 cross-encoder/ms-marco-MiniLM-L6-v2 85.9M downloads a month, the second most-downloaded model on Hugging Face. Old, small, and still extremely hard to beat on cost. | cross-encoder | Apache 2.0 | — | — | 312 | 85.9M |
Qwen3-Reranker-4B Qwen/Qwen3-Reranker-4B The larger reranker option, with a 0.6B sibling for latency-bound paths. Sits after retrieval and before generation. | Alibaba | Apache 2.0 | 4B | — | 155 | 2.4M |
| Time series2 | ||||||
TimesFM 3.0 google/timesfm-3.0-pytorch Zero-shot forecasting from a pretrained model, no per-series training. Trending at 272k downloads within days of release. | Apache 2.0 | 0.3B | — | 565 | 271.7k | |
Chronos-2 amazon/chronos-2 23.9M downloads a month. Foundation models for forecasting have quietly become the default, and this is the most used of them. | Amazon | Apache 2.0 | — | — | 433 | 23.9M |
| Foundation backbones2 | ||||||
bert-base-uncased google-bert/bert-base-uncased 50.7M downloads a month, seven years on. A reminder that most production NLP is not a frontier model and never was. | Apache 2.0 | 110M | — | 3.0k | 50.7M | |
clip-vit-base-patch32 openai/clip-vit-base-patch32 20.5M downloads a month. Still the backbone under a large share of image search, moderation and zero-shot classification. | OpenAI | MIT | — | — | 1.2k | 20.5M |
Every row carries the date we last verified it. Prices are list rates at the standard tier and exclude batch discounts, long-context surcharges and regional variation. Hugging Face download and like counts are pulled from the API rather than retyped, and they measure adoption rather than quality.
Qwen3.8-27B is Apache 2.0 and Qwen3.8-Flash-Next, released three weeks later in the same family, is qwen-community-1.0. GLM-5.3 carries a bespoke Z.ai licence while GLM-5.3-Flash is MIT. FLUX.1-dev is non-commercial and FLUX.1-schnell is Apache 2.0. Assuming a family licence is how a compliance problem gets built into a product.
GLM-5.3-Flash is the most permissively licensed model of its size we are aware of: no field-of-use restriction, no user threshold, no acceptable-use policy that can be revised after you have shipped. It also has more downloads and more likes than its larger, more restrictively licensed flagship, which tells you what practitioners actually optimise for.
The comparison teams make is against GPT-6 Astra at $10 per million. The comparison that matters is against Qwen3.8-Flash at $0.14 or GLM-5.3 at $1.40, because those are the same models served by people whose job it is. Once you count GPU hours, on-call and the engineer maintaining the serving stack, self-hosting usually has to win on privacy, latency or genuine sustained volume rather than on cost.
No vendor classifier between prompt and model, no refusal path, no abuse monitoring. Input filtering, output review, rate limiting and audit logging all become yours to build. This is the most common reason a self-hosted pilot stalls before production, and it is worth scoping at the start rather than discovering at the security review.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-FrameworkDirect API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Real distributions, not averages. We know which routes are slow, and why.
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
