The most important open-weight models in September 2026. Which model fits which , what hardware sits behind it, which licence applies. The figures are checked against the providers' model cards and the library.
Pull models via Ollama
# General purpose with image understanding, Apache-2.0 - runs in about 18 GB at Q4
ollama pull qwen3.8:27b
# Small model for modest hardware, Apache-2.0
ollama pull gemma4:12b
# European alternative from Paris, Apache-2.0
ollama pull mistral-small:24b
# Reasoning and mathematics, MIT licence
ollama pull deepseek-r1:32b
# OpenAI's open-weight model, Apache-2.0
ollama pull gpt-oss:20bEvery model has its own strengths — size, licence, context length, . Here is the compact overview, linked to each provider's official model page. Newer families such as GLM 5.3, DeepSeek V4.1 or MiniCPM5 already exist on Hugging Face but were not yet in the library on 11 September 2026, which is why they are not listed here.
Meta
Meta flagship
Llama 4 comes as a mixture of experts: Scout with 16 experts, Maverick with 128, each with 17 billion active parameters. Meta has not released a new generation since April 2025. The Llama licence permits commercial use with conditions.
ollama pull llama4:scout
Mistral AI
French powerhouse
Mistral Small 3 with 24 billion parameters is the practical size for a single graphics card. Above it sit Mistral Small 4 and Mistral Large 3 for server hardware. All Apache-2.0, and the provider is based in Paris.
ollama pull mistral-small:24b
Alibaba
The most downloaded open-weight candidate
Qwen 3.8 with 27.3 billion parameters also understands images and holds around 262,000 tokens of context. Apache-2.0, and therefore unproblematic for mid-sized businesses. Excellent German quality — this is the model we use day to day ourselves.
ollama pull qwen3.8:27b
Google, now under Apache-2.0
4 ranges from the tiny E2B and E4B variants up to 31 billion parameters, plus a mixture-of-experts version with 26 billion. The key change from 3: the separate licence is gone, 4 is Apache-2.0.
ollama pull gemma4:12b
DeepSeek
Reasoning specialist
DeepSeek R1 still sets the standard for logical reasoning and mathematics. The full version with 671 billion parameters needs server hardware; the distilled variants with 32 and 70 billion run on hardware a mid-sized company actually buys. MIT licence.
ollama pull deepseek-r1:32b
Microsoft
Microsoft small champion
Phi 4 with only 14B parameters shows: well-curated training data beats sheer size. Very good performance on a single RTX 4090. MIT license, commercially unrestricted.
ollama pull phi4:14b
OpenAI
OpenAI's open-weight model
In August 2025 OpenAI released gpt-oss in two sizes (20B and 120B) under Apache-2.0. In model quality clearly better than older OSS models, with full commercial release.
ollama pull gpt-oss:20b
An open-weight model is an whose trained weights (the parameter file) are publicly available — typically as a download on HuggingFace, on the vendor website or through . That distinguishes them from ChatGPT, Claude or Gemini Pro, whose weights are not released and which are accessible only via APIs.
Important: “open weight” is not the same as “open source”. Licences vary widely — from genuine MIT or Apache-2.0 (Qwen, 4, Mistral, Phi, DeepSeek, gpt-oss) to bespoke provider licences with conditions (Llama). Anyone deploying commercially reads the licence before the download, not after.
Recommendations from our own client work, as of September 2026. Not “the best model” — there is no best. Rather: for this purpose, this model is the most pragmatic fit.
Six properties that make these seven models a class of their own — and distinguish them from cloud frontier models.
OpenAI-compatible API call against a local model
# Pick a model (variable)
MODEL="qwen3.8:27b"
# Alternative: gemma4:12b · mistral-small:24b · deepseek-r1:32b · gpt-oss:20b
# OpenAI-compatible chat-completion call
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "'$MODEL'",
"messages": [
{"role": "system", "content": "You are a pragmatic AI consultant for SMBs."},
{"role": "user", "content": "Which model do you recommend for an 8-person law firm with an RTX 4090?"}
],
"temperature": 0.2
}'
# The same request also works with the OpenAI Python SDK,
# simply set OPENAI_API_BASE to http://localhost:11434/v1.Cloud alternatives honestly compared
Three cloud frontier models that are still often ahead in absolute quality. Trade-off: data flow and per-token costs against top quality.
Frontier cloud (USA)
OpenAI
Frontier cloud (USA/EU)
Anthropic
Frontier cloud (USA)
Rule of thumb: with GDPR obligations or IP protection — open-weight local. For maximum quality on uncritical content — cloud frontier. For both — multi-routing in (sensitive content local, general research cloud).
Pricing
License
Mix of Apache-2.0 (Mistral, Qwen, Phi, GPT-oss), MIT (DeepSeek), custom community licences (Llama, Gemma). For SMB own-use in nearly all constellations unproblematic — verify per model on the respective card.
Hardware costs
Mid-range (for models up to 32B): RTX 4090 24GB + 32GB RAM workstation, around €3,000. Premium (for 70B+): Apple Mac Studio M4 Ultra 192 GB from €8,000 or an NVIDIA server with RTX 5090. Power: 150–300 W during inference, <50 W idle.
vs. cloud API
Cloud costs scale linearly with volume. At 1,000 tokens/day per 10 staff: GPT-5 around €60/month, Claude Sonnet similar. Self-hosted: hardware amortisation in 12–18 months. From 50+ staff clearly cheaper.
Important: quantisation quality is not linear. Q4_K_M is standard and 'good enough' for 95 % of . Q5/Q6 for higher demands. Full FP16 only for research. Anyone trying to make do with less VRAM should first check quantisation before picking a smaller model.
Related topics
Models alone are useless. loads them, provides the user surface, the server solution sits behind:
Free intro call, no strings attached. In 30 minutes you'll know whether and how AI can help your business.