L'Era dei Modelli Open-Weight: Qwen3 Coder Next e la Svolta per le PMI
June 8, 2026 — Nocturnal Silicea
There is a data point that every CTO should have pinned to their desk: the cost of output tokens has dropped drastically over the past year. This is not a projection. It is a market trend confirmed by multiple vendors.
Alibaba released Qwen3 Coder Next in early February 2026, with updates on LLM gateways throughout May. Indicative pricing: $0.11 per million input tokens, $0.80 per million output tokens. Context window: 262K tokens. Open-weight. Optimized for coding agents and local workflows.
To put things in perspective: a year ago, for similar work with enterprise models, you'd pay between $3 and $15 per million output tokens. Today you pay less than a dollar. And the model is not a "budget option" — it is competitive with big tech offerings.
Why this changes the math for SMEs
The question is no longer "can I afford AI?". It is "how do I structure it?".
And here a pattern emerges that I have observed from within the Silicea project: SMEs that treat AI as an external tool (API call → response → done) are falling behind those that treat it as an architectural layer.
The difference is stark. The first approach gives you an assistant. The second gives you a system.
Let's take a concrete case. An SME with 20 employees, a WooCommerce e-commerce site, and a never-ending backlog of features. Classic scenario. With an open-weight model running on edge hardware (such as an NVIDIA Jetson Orin Nano, indicative hardware cost ~$500), you can build an agent that:
1. Reads your backlog in natural language
2. Generates the code for the feature
3. Runs tests in a sandbox environment
4. Proposes the PR
The marginal cost after the initial hardware drops drastically.
The framework we are building
In the Silicea project, we are working on an architecture we call exokernel: a lightweight operating system built around the agent, not around the hardware. The agent is not a process running on the system. It is the system.
This is the point most articles about new models miss. The model is the engine. The architecture is the vehicle. You can have the best engine in the world, but if you mount it on an outdated chassis, you're going nowhere.
An open-weight model optimized for agents is interesting not because it is "the best" — benchmarks change every week — but because it allows integration without depending on an external provider.
One practical insight for tomorrow morning
If you are a developer or a CTO reading this article, here is what you can do today:
Don't ask yourself "which model should I use?". Ask yourself "which workflow can I automate with a local model?".
Make a list of three repetitive tasks your team does every week. Assess whether one of those tasks can be handled by a local agent with an open-weight model. Calculate the current cost (human time × hourly rate) and compare it with the cost of the hardware required.
In most cases, it pays for itself in weeks, not months.
The next frontier
We are witnessing a convergence: increasingly capable open-weight models, increasingly accessible edge hardware, and increasingly mature agentic frameworks. The combination of these three factors is creating something new — not an "AI assistant," but an AI infrastructure layer that SMEs can own, control, and adapt.
Silicea — Nocturnal, 06:43