A 1.17-billion-parameter model scored 85.5% in its creator’s tests, while Qwen2.5-1.5B managed 53.3%. The real story is not “yet another small model,” but the possibility that a single RTX 5090 and roughly 11 billion training tokens could create a practical advantage. However, the author supplied the data, code and model, so the conclusions still require external validation.
What this is
ConeML is a sparse mixture-of-experts model (MoE, an approach that activates only some parameters for each input to reduce computing costs). The model stores about 1.17 billion parameters, but activates roughly 348 million per token, and its developer trained it on a single RTX 5090. The Q8 quantized version is about 1.25GB and can run on a CPU.
The author designed a “fact-chain” test that shuffles the relative relationships between people or objects and asks the model to follow those relationships to an endpoint. The longer the chain, the harder the task. ConeML outperformed Qwen2.5-1.5B, SmolLM2-1.7B, Llama 3.2 1B and Gemma 3 1B at chain depths of 1, 5, 12 and 24.
Industry view
We believe this shows that model efficiency may depend on more than parameter count and training data volume. Architecture, training methods and task-specific design can all change the cost-performance equation. For enterprises and small teams seeking local deployment with limited computing budgets, this quantifiable model, which can run on ordinary hardware, deserves a place on the testing shortlist.
But the objections are equally clear: the author acknowledges that ConeML achieved only 13.8% accuracy on a task comparing relationships within chains and scored just 12/164 on the HumanEval programming test. It also failed all six two-turn dialogue tests. It looks more like a model that has mastered a specific lookup strategy than one with reliably general understanding. Moreover, the 85.5% score came from a test built by the author, and the Q8 version has not yet been shown to reproduce the FP32 results. The model therefore cannot yet be considered superior to mature alternatives.
Impact on regular people
For enterprise IT: Small models are continuing to lower the deployment cost of document retrieval, knowledge organization and fixed-process tasks. Before formal adoption, teams should retest accuracy, speed and hardware compatibility.
For individual professionals: Local execution, quantization and single-GPU training give more teams outside large companies the conditions to build custom models, increasing the value of combining domain expertise with engineering skills.
For the consumer market: Smaller, faster local models could reach personal computers and edge devices. But until their capabilities in general conversation and complex programming are demonstrated, they are better suited to specialized tools in the near term than to replacing general-purpose assistants.