The first checkpoint won't be released. After a local fine-tuning round on AliceAI-80B-A3B-Base, the project's creator decided not to publish version 1.0 alpha: it technically works, but it produces garbled output as soon as a request moves away from the examples seen in training. This is not a hardware failure; it is a lesson about data.

The project uses three 32GB V100s to train an instruct fine-tune of AliceAI-80B-A3B-Base. To teach chain-of-thought reasoning and conversation, the author uses a shallow distill of Qwen 3.8 27B. All training data is generated locally on the same GPUs via sftmill, an off-policy distillation engine released as open source. The first dataset of 5 million tokens was oriented toward coding; the model learned chain-of-thought reasoning and conversational responses, but it did not cover enough ambiguous requests.

The bottleneck is not VRAM, it is data distribution

This is the core point. Five million tokens might seem like a substantial volume for fine-tuning, but if the corpus is skewed toward a specific domain, the model does not learn to handle the noise and variation typical of real instructions. The result is selective underfitting: prompts close to the examples work, while more distant ones produce disconnected text. Checkpoint 1.0 alpha is not broken; it is simply too trained on a niche.

More breadth, not more tokens

The answer is not to change GPUs but to change the composition of the data. The creator increased the synthetic generator from 80 to about 240 total tokens per second, distributing the work across three instances with four parallel workers each. The new dataset, still about 5 million tokens, will be more oriented toward general instructions, and training will resume from checkpoint 1.0 alpha with a reduced learning rate.

There is a structural signal here for people working with on-premise LLMs. In local environments, the advantage is not just privacy and control: it is also the ability to iterate on data without going through external APIs. The downside is the synthetic generation bottleneck: if a pipeline produces 80 tokens per second, a large dataset takes a long time; at 240 tokens per second cycles shrink, but the limiting factor remains prompt variety. For those evaluating on-premise deployments, the trade-offs among older hardware, iteration speed, and data control are exactly the kind of territory covered by AI-RADAR's analytical frameworks at /llm-onpremise.

The second round will show whether the broader-data strategy works. In the meantime, the sftmill repository remains public and the creator has promised to share the first actually usable checkpoint. Anyone with three V100s and a lot of patience can follow the same path: the real luxury is not the model, it is control over the pipeline.