Open-Weight Models
Open weights turned the best AI into something you can own.
Open-weight model releases changed the enterprise math: frontier-adjacent capability you can download, fine-tune, and run behind your firewall forever. NetRay turns that possibility into production: license review, honest benchmarking on your tasks, and serving infrastructure that holds up.
What open weights buy you (and what they don't)
Open weights mean the model file is yours to run: no per-token meter, no deprecation on a vendor's schedule, no terms-of-service change pulling a capability your product depends on. They also mean the operational responsibility is yours, which is where most teams stall and where we do our work.
What they don't automatically buy: frontier performance on open-ended reasoning (hosted flagships still lead there) or production readiness out of the box. The winning enterprise pattern is narrower and better: a mid-size open model, fine-tuned to your domain, serving your specific workloads faster and cheaper than any API.
Choosing among the families
Llama-class models bring the broadest ecosystem and tooling; Mistral-class models punch above their weight at smaller sizes; Qwen-class models lead many multilingual and coding benchmarks; and the distilled/specialist variants beneath each family are often the real production picks. This landscape shifts quarterly, which is why we maintain a standing evaluation harness rather than a standing opinion.
Our selection method is stubbornly boring: define your evaluation set first, benchmark candidates on your tasks and your hardware, and pick the smallest model that clears the bar. Every point of unnecessary size is latency and hardware you pay for forever.
Licensing: read before you build
Open-weight is not one license. Some families are permissive for nearly all commercial use; some carry acceptable-use policies, user thresholds, or restrictions that matter to specific industries; and fine-tuned community derivatives inherit terms from their base. We run license review as a formal step, with your counsel where needed, before a model enters your registry. Boring, and cheaper than discovering a restriction after the product ships.
From download to dependable
The path we run: evaluation set, benchmark, license review, quantization to your hardware budget, fine-tuning where the numbers justify it, then production serving with the full observability stack. When the next major release lands (they keep coming), your harness answers 'should we switch' in 48 hours with numbers instead of a month of debate.
Straight answers
Are open models safe for enterprise use?
Operationally, yes, with the same engineering any production dependency gets: pinned versions, signed artifacts, security review of the serving stack, and output guardrails tuned to your policies. The model file itself is inert weights; the risk surface is the serving infrastructure, which you control.
How far behind hosted frontier models are they?
On broad reasoning benchmarks, months, and the gap keeps cycling as releases leapfrog. On your specific workload after fine-tuning, the 'gap' frequently inverts: tuned open models win narrow tasks outright. The only answer that matters is your evaluation set's answer.
Can we redistribute or embed a fine-tuned model in our product?
License-dependent, and this is precisely why review comes first. Many licenses permit it cleanly; some require attribution or restrict certain uses. For products (as opposed to internal tools) we map the license to your distribution model before training begins.
What happens when better models are released?
You upgrade on your schedule, or don't. That's the ownership difference: a hosted API can deprecate your model out from under you, while a model in your registry runs until you choose to replace it, and your evaluation harness makes replacement a data decision.
Benchmark the open-model landscape on your tasks.
We'll evaluate the current leading open-weight candidates against your workload and hardware, and hand you the table.
Request a model evaluationFree AI workshops available. A founder reads every message.