When a model answers '2' to 'What is 1+1?' without adding an emoji or a line of explanation, it is not just a stylistic curiosity. Oc-e-Non-v1, built from Qwen3-4B-Instruct-2507, cuts average response length in half and brings into focus an often overlooked aspect of local inference: the cost of output tokens.

The name refers to the nickname some contemporaries reportedly gave Richard I of England, Oc-e-Non, meaning 'yes and no' in Occitan, because of his tendency to use few words. The project's author used Heretic, a framework for modifying LLM behavior, together with the config.terse.toml configuration file. This is not a simple system prompt: concise behavior is baked into the model, a meaningful difference for anyone seeking predictable results without relying on textual instructions the model can interpret inconsistently.

The base model is Qwen3-4B-Instruct-2507, described as the last native non-thinking Qwen version the author is aware of. Reasoning effort control is not yet available in Heretic, which makes this model a useful testbed for separating conciseness from reasoning capability.

Beneath the light tone there is a concrete constraint. Text generation is sequential: each additional token requires a decoding step. In a self-hosted deployment, shorter responses improve not only readability; they reduce time to completion and, on batch workloads or queued requests, free up compute capacity. On a model with only a few billion parameters, the absolute gain per call may look modest, but across thousands of requests the effect on throughput is structural.

Forced conciseness is not neutral, however. Removing explanations and intermediate steps can reduce transparency and verifiability. In regulated contexts or pipelines where a decision must be justified, a dry answer like 'Yes' can be a problem rather than an advantage. The value depends on the workload: software agents that need to pass compact outputs to another module, data extraction systems, and cost-sensitive APIs benefit from tersification; educational assistants, debugging tools, and analyses requiring reasoning may lose useful information.

For anyone evaluating on-premise deployment, experiments like this signal that output efficiency is becoming a design lever alongside quantization, caching, and hardware choice. If a behavior can be embedded in the model, the role of prompt engineering also changes, from continuous correction to initial configuration.

The author invites readers not to take the project too seriously, calling it a light-hearted experiment while waiting for Qwen4. But the open question it leaves is concrete: how much verbosity can be removed before an answer stops being useful?