Open-source and open-weight AI models
Open source, open weights and open access are not interchangeable. Before adopting a model, inspect its licence, deployment method, provider and data path.
THE SHORT ANSWER
Choose by task, not by logo.
GPT-OSS, Qwen and DeepSeek are useful candidates for an open-model evaluation. Compare capability and cost, then separately verify licensing, hosting control and data-handling requirements.
GPT-OSS 120B
OpenAI's open-weight reasoning model, routed by OpenRouter.
- Provider
- OpenRouter · OpenAI
- Prima Ordia cost
- 0 credits
- Images
- Text only
Qwen 3.6 Beta
Multilingual reasoning and coding from Alibaba's Qwen family.
- Provider
- OpenRouter · Alibaba Qwen
- Prima Ordia cost
- 0 credits
- Images
- Text only
DeepSeek V4 Flash
Cost-efficient reasoning and coding.
- Provider
- DeepSeek
- Prima Ordia cost
- 1 credit
- Images
- Text only
01 / WHERE IT FITS
Good reasons to test these models
- Evaluating open-model capability
- Efficient reasoning and coding
- Avoiding one closed-model dependency
- Planning a future self-hosted option
02 / KEEP YOUR GUARD UP
What a useful comparison must catch
- Remote access is not self-hosting
- Read the exact licence
- Infrastructure costs can outweigh token savings
- Privacy depends on the complete deployment
03 / THE SCORECARD
Judge the finished work, not the demo.
Give every model the same context and constraints. Score each dimension from one to five, then include the time you spent correcting the answer.
Answer quality
Does the answer solve the task accurately, completely and at the right level of detail?
Reasoning
Can it handle constraints, expose assumptions and recover when the first approach fails?
Speed
Is it responsive enough for the way you actually work, including revisions?
Value
Does the result justify its credit cost for this particular task?
DON'T CHOOSE BASED ON OUR OPINION
Test them yourself.
04 / A FAIR TEST
Five rules that make the result worth trusting
- Use real work.Choose a task you repeat, not a trick question designed for a leaderboard.
- Hold the prompt constant.Same context, constraints, requested format and deadline for every model.
- Define success first.Write down what a correct, useful answer must contain before you see any output.
- Run more than one example.Include an easy, typical and difficult case so one lucky result cannot decide.
- Count correction time.The cheapest or fastest response loses if it creates more work before acceptance.
05 / COMMON QUESTIONS
What people ask
Is an open-weight model open source?+
Not automatically. Source code, training data, weights and licence permissions are separate questions. Read the model's current licence.
Does using an open model keep data private?+
Only if the complete deployment provides that control. A hosted model still processes prompts on remote infrastructure.
Why test open models in Prima Ordia?+
It provides a convenient way to compare configured hosted versions before considering a separate deployment decision.