PrismML: Ternary Bonsai 2 27B

Explain how RLHF training works to a software engineer who understands backpropagation but hasn't worked with reinforcement learning before.

RLHF adds a preference-learning layer on top of standard supervised fine-tuning. You start with a base model trained on next-token prediction. Then collect human rankings: show evaluators multiple model outputs for the same prompt and ask which is better. Train a reward model (usually a classifier) to predict these human preferences — it learns to score outputs the way your evaluators would. Now the key part: use that reward model as a differentiable objective. Generate outputs from your language model, score them with the reward model, and backprop the reward signal to update the language model's weights. This is typically done with PPO (Proximal Policy Optimization), which prevents the model from drifting too far from its original behavior in any single update. The result: a model that optimizes for human-judged quality rather than just likelihood under the training distribution.

This example highlights the model's technical explanation ability — it correctly sequences concepts and bridges from familiar ground (backprop) to new territory (RL). The explanation is accurate and appropriately scoped for the stated audience. With multimodal support (text + image), this model could theoretically include diagrams in explanations, though we haven't demonstrated that here. The lack of public benchmarks means we can't compare its MMLU or reasoning scores to established models.

When 262K context beats chunking for contract review teams

A 4-person legal ops team processing M&A due diligence needs to cross-reference clauses across 40+ contracts in a single pass. Ternary Bonsai 2 27B's 262,144-token window fits roughly 200,000 words—enough to load an entire deal room and ask questions that span documents without retrieval chunking. At $0.07 input per million tokens, a 200K-token load costs $0.014, making exploratory queries cheap even when you're iterating on search terms. The $0.50/Mtok output rate means a 2,000-token summary runs $0.001. If your team runs fewer than 100 discovery sessions per month and needs genuine cross-document reasoning (not keyword search), this model's context advantage justifies the lack of public benchmarks. Beyond that volume, test against Claude 3.5 Sonnet for accuracy on nuanced clause interpretation.

Multimodal QA for e-commerce teams with visual SKU data

A 3-person Shopify agency audits product listings where images carry critical spec details (fabric swatches, dimension diagrams, ingredient labels). Ternary Bonsai 2 27B's image + text modality lets you feed a product photo and description together, then ask "Does this image match the listed material composition?" or "Flag dimension mismatches between the diagram and the text." At $0.07 input, processing 1,000 image-text pairs costs roughly $70 in tokens (assuming ~1K tokens per image encoding), which pencils out if you're catching listing errors that cost more in returns. The model's 27B parameter count suggests it won't match GPT-4V on complex visual reasoning, but for structured QA where the image is a reference document (not a scene to interpret), the price-per-check is competitive. Run a 50-item pilot before committing to bulk workflows.

Why this model loses to faster options for live support routing

A 10-person SaaS support team wants to auto-route inbound chats based on message content and urgency. Ternary Bonsai 2 27B's $0.50/Mtok output rate looks cheap until you calculate throughput: a 200-token routing decision on 500 chats/day costs $50/day in output tokens alone, and the 27B parameter count implies slower inference than 7B-class models optimized for latency. For real-time triage where response time under 500ms matters, a smaller model like Llama 3.1 8B at $0.05 input / $0.08 output delivers sub-200ms p95 and cuts your monthly token bill by 80%. Ternary Bonsai 2 27B makes sense only if your routing logic requires the full 262K context (e.g., loading a week of chat history per decision), which is rare in live support. For standard triage, pick speed and cost over context depth.