A year ago, choosing an open model mostly meant picking a Llama size. In 2026 the field is crowded with trillion-parameter open-weight models from China, a newly Apache-licensed Gemma from Google, Meta stepping back from open releases, and small models good enough to run on a phone. For businesses that want privacy, control or lower costs, there has never been more choice — or more fine print.
A note on terms: most of these are open-weight — you can download and run them — but many come with custom licences that are not "open source" in the strict sense. We flag the licence for every model below.
Who is leading? Independent rankings
Rankings move weekly, so treat any snapshot as approximate. On Artificial Analysis's open-weights leaderboard (Intelligence Index v4.3, checked October 14, 2026), the top open models were Xiaomi's MiMo-V2.6-Pro, Zhipu's GLM-5.3, Moonshot's Kimi K3, followed by GLM-5.3-Flash, DeepSeek V4.1 Flash and Alibaba's Qwen3.8 27B. Almost all the leaders are Chinese labs; most benchmark scores beyond these rankings are self-reported by vendors.
The leading open LLMs in 2026
DeepSeek V4 family
DeepSeek's V4 line made a one-million-token context the default and ships under the permissive MIT licence for both code and weights. DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active per token, and DeepSeek says it needs a fraction of the compute and memory of V3.2 at long context (model card). The lighter DeepSeek-V4.1-Flash activates only 8–16 billion parameters, accepts images as well as text, and is also MIT-licensed (model card).
Licence: MIT — among the most permissive at this scale.
Qwen 3.8 (Alibaba)
Qwen remains the most complete open family. Qwen3.8-27B is a dense 27B model under Apache 2.0 with native image and video understanding and a 262K context extensible to one million tokens (model card) — an excellent single-GPU workhorse. The giant Qwen3.8-2.4T-A95B "Max" model is open-weight but uses a custom Qwen licence, not Apache (model card). Alibaba says Qwen 4 is currently in training, with no release date.
Licence: Apache 2.0 for 27B; custom for the largest models — check each one.
Kimi K3 (Moonshot AI)
Kimi K3 is one of the strongest open models available: 2.8 trillion total parameters with 104 billion active and a one-million-token context (GitHub). Its weights are about 1.4 TB, so it is a data-centre model.
Licence: a custom Kimi K3 licence. Internal use is unrestricted, but companies offering it as a hosted service with more than $20 million in revenue need a separate agreement, and very large consumer products must display "Kimi K3" (licence).
GLM-5.3 (Zhipu / Z.ai)
GLM-5.3 is a 753B-parameter model that Z.ai positions as its most capable open model for coding (model card). Earlier GLM-5.x releases are MIT-licensed; GLM-5.3 moved to a custom licence that remains free for commercial use but requires a security review for hosting providers above $10 billion in revenue.
Xiaomi MiMo-V2.6
Xiaomi's MiMo-V2.6-Pro, a one-trillion-parameter model with 42 billion active and a one-million-token context, topped the Artificial Analysis open-weights index in October 2026 and is reported to be MIT-licensed.
Google Gemma 4
Gemma 4 is the big licensing story of the year: Google released it under Apache 2.0, replacing its earlier custom Gemma terms (Google). The family runs from E2B and E4B models that work offline on phones and a Raspberry Pi, through a 12B model, a 26B mixture-of-experts model with under 4B active parameters, to a 31B dense model. All sizes understand images, the smaller ones also take audio, and the 26B and 31B models fit on a single 80 GB GPU (model card). It was also the most-used open model family in production in a September 2026 enterprise survey.
Licence: Apache 2.0.
Mistral
Mistral's open line-up includes Mistral Small 4 (Apache 2.0), which merged its reasoning, vision and coding models into one, and Mistral Large 3 (Apache 2.0). Mistral Large 4, a model of more than one trillion parameters, launched as a preview API in October with weights promised by the end of the month; its licence had not been published at the time of writing (Mistral docs). Mistral Medium 3.5 uses a modified MIT licence with carve-outs for large companies.
Meta: from Llama to Muse
Meta's flagship Muse Spark (April 2026) is proprietary, a sharp turn away from open Llama releases (VentureBeat). In August Meta released Muse Glimmer, a 30B model distilled from Spark under Apache 2.0, small enough to run in under 16 GB at 4-bit (The Register). Llama 4 (2025) remains available under its community licence.
Others worth knowing
- OpenAI gpt-oss-120b and gpt-oss-20b — Apache 2.0, released August 2025; no successor yet (OpenAI).
- NVIDIA Nemotron 3 — efficient hybrid models with a one-million-token context; the 550B Ultra needs a multi-GPU server (model card).
- IBM Granite 4.1 — 3B, 8B and 30B models under Apache 2.0 with up to 512K context, built for enterprise use (IBM).
- Microsoft Phi-4 — small MIT-licensed models, including a 15B vision-reasoning model.
- MiniMax-M3 — strong multimodal model, but its community licence requires attribution and written authorisation for companies above $20 million revenue (licence).
- Ai2 OLMo 3 — fully open: data, code and checkpoints, ideal for research and auditability.
Licence cheat sheet
| Model | Licence | Commercial use |
|---|---|---|
| Gemma 4, Qwen3.8-27B, Mistral Small 4, Granite 4.1, Muse Glimmer, gpt-oss | Apache 2.0 | Yes |
| DeepSeek V4 / V4.1, Phi-4, GLM-5 to 5.2 | MIT | Yes |
| Kimi K3 | Custom | Yes, with conditions for large hosting providers |
| GLM-5.3 | Custom | Yes, with review for very large hosting providers |
| MiniMax-M3 | Community licence | Attribution; authorisation above $20M revenue |
| Llama 4 | Community licence | Yes, unless over 700M monthly users |
| Mistral Medium 3.5 | Modified MIT | Carve-outs for large companies |
Always read the current licence on the model card before deploying — terms change between versions.
Which open model should you use?
- Single GPU or on-premise server: Gemma 4 31B or 26B, Qwen3.8-27B, Muse Glimmer or Granite 4.1 30B.
- Laptop, phone or edge device: Gemma 4 E2B/E4B or small Phi and Granite models.
- Maximum capability, self-hosted cluster: DeepSeek V4-Pro, Kimi K3, GLM-5.3 or MiMo-V2.6-Pro.
- Coding: GLM-5.3, Qwen's coder models and Kimi K3 are strong; compare on your own repositories.
- Strictest licensing requirements: stick to Apache 2.0 or MIT models.
- Auditability and research: OLMo 3, with fully open data and training code.
Are businesses actually using open models?
Increasingly, though surveys disagree on how much. An Enterprise Technology Research survey reported in September 2026 found open-weight models handle 34% of enterprise AI tokens, up from 23% a year earlier, with 42% of respondents running them in production — and cost savings the top reason (Techstrong.ai). Menlo Ventures, measuring differently, found open models' share of enterprise LLM usage fell in 2025 (Menlo Ventures). The likely reality: companies use closed frontier models for the hardest tasks and open models where privacy, cost or control matter most.
Open vs closed: the trade-offs
| Open-weight models | Closed API models | |
|---|---|---|
| Data control | Stays in your infrastructure | Sent to the provider (under contract) |
| Cost at high volume | Often lower | Pay per token |
| Upfront effort | GPUs, deployment, security | Minimal |
| Customisation | Full fine-tuning | Limited |
| Frontier capability | Close, but usually behind | Usually ahead |
| Vendor risk | Low — you keep the weights | Model changes and retirements |
For the practical side of running these models, read our guide to self-hosting LLMs. For vision and documents, see open-source vision-language models and the best open-source OCR models.
How Zinetics can help
We help companies choose, deploy and integrate open models — from a private assistant running on your own servers to a hybrid setup that routes sensitive data to a self-hosted model and complex tasks to a frontier API. Explore our AI services or talk to us about your requirements.