Open-weight models in 2026: close to the frontier, not always as open as they look

Open weights are not open source

A model has three parts: the weights (the trained parameters you download), the code to train and run it, and the data it learned from. "Open-weight" only promises the first, under a license that may or may not be permissive. The leading 2026 releases from DeepSeek, Qwen, Kimi or GLM publish weights and a technical report, not their training data.

Few projects publish everything: Ai2's Olmo 3 and 3.1 (weights, code, checkpoints and datasets under Apache 2.0) and the Swiss AI Initiative's Apertus v1.5, which respects data owners' opt-outs and ships EU AI Act documentation.

Why it matters: without the data you cannot audit what the model learned, and the European Data Protection Board warned in December 2024 that unlawfully processed training data could affect the lawfulness of a deployment. The realistic answer is to document provenance and test outputs.

The main families in September 2026

Sizes, context and license, from each vendor's model card.

The flagships reach 1 to 2.8 trillion parameters and target large servers: Moonshot recommends 64 or more accelerators for K3. The efficient middle is MoE with few active parameters (DeepSeek V4.1-Flash, GLM-5.3-Flash, Mistral Small 4, gpt-oss-120b). And a dense class of 27B to 31B runs on one 32 GB GPU once quantized, as our local AI hardware guide shows.

Strengths by use: code, reasoning, languages, vision

• Coding and agents. GLM-5.3 is the best open-weight entry on the official Terminal-Bench 4.0 board (41.8%, fifth overall). On Terminal-Bench 2.1, with vendor-reported figures, DeepSeek V4.1-Flash (90.6), MiMo-V2.6-Pro (89.9) and Kimi K3 (88.3) lead; Qwen3.8-27B reports 61.7 on SWE-bench Pro.
• Reasoning. Vendors report up to 93.5 on GPQA Diamond (Kimi K3), but the test is saturated (the best closed models sit at 94 to 96%) and no longer separates models.
• Spanish and French. The Qwen3.5 generation covered 201 languages; Gemma 4 supports 35+ out of the box; Mistral Large 3 lists 40+ native languages, and Cohere's Command A+ (Apache 2.0) names both among 48. Regional options such as Latam-GPT, Spain's ALIA-40b or France's Luciole deserve a test for language-heavy work.
• Vision and audio. Qwen3.8-27B, Gemma 4, Kimi K3, MiniMax-M3, GLM-5.3-Flash and DeepSeek V4.1-Flash accept images; MiMo-V2.6 also takes audio and video.

How far behind the frontier

Independent leaderboards, read on September 25 and 26, 2026.

Epoch AI measured an average lag of about four months for the best open-weight models since January 2026. The gap is small on chat preference (Arena) and wider on long agentic tasks (Terminal-Bench), exactly where frontier models earn their price. Among small models, Qwen3.8-27B scores 34 on the Artificial Analysis index, far above Gemma 4 31B (19) and Muse Glimmer (17).

The license drift, and how to check a license

In 2025 most leading open weights shipped under MIT or Apache 2.0. In 2026 several flagships moved to custom licenses, while smaller siblings often stayed permissive (Qwen3.8-27B is Apache 2.0, GLM-5.3-Flash MIT): free for most companies, with conditions that bite at scale or on certain business models.

• Kimi K3 (K2 was Modified MIT): name display above 100 million monthly users or $20 million monthly revenue, and an agreement for model-as-a-service businesses above $20 million in 12 months.
• Qwen3.8 flagship: the same display rule and a separate license for model-as-a-service or AI work-assistant businesses above $50 million in group revenue. Qwen3.8-Flash-Next requires it for such businesses at any size.
• GLM-5.3 (GLM-5.2 was MIT): a Z.ai security review for model-as-a-service operators with group revenue above $10 billion.
• MiniMax-M3: "Built with MiniMax M3" attribution, a notice or prior authorization depending on a $20 million yearly revenue threshold, and banned uses such as military; M2.7 was non-commercial.
• Hy3-preview (Tencent) excluded the EU, the United Kingdom and South Korea; the final Hy3 is Apache 2.0.

1. Read the LICENSE file of the exact version, not the metadata: MiniMax-M2.5's card said "modified-mit" while its file added banned uses and Singapore law.
2. Check commercial use and user or revenue thresholds, including group revenue.
3. Check model-as-a-service and work-assistant clauses if you resell access or build an assistant.
4. Note attribution, territory, banned uses and governing law.
5. Store license, version and file hash in your AI bill of materials and re-check at every upgrade.

Where to run them, and where your data goes

Open weights give you three hosting options, each with a different data flow. Self-hosting keeps prompts and outputs inside your perimeter: a Chinese model on your server sends nothing to China, because weights are just files. Inference providers run the same weights under their own contract and jurisdiction. The lab's API sends your data to the lab.

The middle has become cheap. On September 26, 2026, gpt-oss-120b cost $0.15 / $0.60 per million input and output tokens on Groq, Together or Fireworks; DeepSeek V4.1-Flash, $0.30 / $1.20 on DeepSeek's API at peak hours and on Fireworks; Qwen3.8-27B, $0.80 / $4.00 on Groq. In Europe, Mistral's platform also lists Z.ai's GLM-5.3 and GLM-5.2: a Chinese-origin model under a French provider's contract; check the hosting region.

The legal test is where personal data goes. In Colombia, Law 1581 allows transfers to countries on the SIC's adequacy list, which includes the United States but not China; for other destinations you need an exception, such as the data subject's express authorization. In the EU, transfers outside the EEA fall under Chapter V of the GDPR.

Risks: provenance, safety tuning and model files

Provenance and geopolitics. Menlo Ventures estimated in December 2025 that Chinese open models accounted for about 1% of LLM API usage among US enterprises. The US defense authorization act for fiscal 2026 requires removing DeepSeek from Defense Department systems, and Zhipu (Z.ai) has been on the US Entity List since January 2025. If you sell to governments or regulated sectors, ask about model origin even when you self-host.

Safety tuning is thin. Research shows that an open model's refusals can be erased by editing its weights, and that the edit works across languages; the Financial Times reported Gemma 4's guardrails stripped within 90 minutes of release. Add a guard model (Llama Guard 4 or gpt-oss-safeguard) and test it in Spanish and French: see guardrails in production.

Model files are software. Pickle checkpoints can run code when loaded, JFrog found about 100 malicious models on Hugging Face in 2024, and instructions can hide in a GGUF chat template. Prefer safetensors from official organizations, pin hashes and keep servers private: SentinelLABS and Censys counted 175,108 exposed Ollama hosts. More in AI supply chain security.

The upside: when commercial models refused forensic requests during a July 2026 intrusion, Hugging Face analysed the attacker's logs with the open-weight GLM-5.2 on its own servers.

How to decide, in Colombia and in Europe

1. Name the dominant constraint: quality, cost per task, latency, data residency or offline use.
2. Shortlist three models that fit your license needs and hardware.
3. Evaluate them on at least 30 real prompts in your languages, against one closed model.
4. Self-host the model for personal or confidential data; use a contracted provider for the rest.
5. Record license, version and hashes, and add a guard model.
6. Plan upgrades: models and licenses change every few months.

Colombia and Latin America. Open weights are the cheapest route to data residency, and the SIC's 2024 guidance on AI requires a privacy impact assessment for high-risk uses. The practical path: a self-hosted open model for sensitive flows and a closed API for the hardest tasks.

European Union. The AI Act has applied generally since August 2, 2026. Under the Commission's guidelines, fine-tuning an open model makes you a general-purpose AI provider only if you use more than a third of the original training compute.

Our recommendation: test Qwen3.8-27B or Gemma 4 31B self-hosted for internal work and DeepSeek V4.1-Flash through a contracted provider for volume, and keep a closed frontier model where the gap shows. If a client excludes Chinese-origin models, Gemma 4, Muse Glimmer, gpt-oss and Mistral are the Apache 2.0 alternatives.