Reflection 70B

Reflection-Tuning is the technique that gives the model its name. Instead of only training on final answers, it trains the model to review its own reasoning and revise it mid-generation. In practice the model produces a draft of its thinking, checks it, and rewrites the parts it judges wrong before delivering an answer. That ordering is the whole idea: correct, then commit.

The model separates its internal work from its answer using special tokens. Reasoning goes inside <thinking> tags, and the final response goes inside <output> tags. This split keeps the scratch work visible, which is handy when you want to debug why a model reached a conclusion, or when you'd rather hide the thinking and show only the answer.

When the model notices a mistake, it emits a <reflection> block inside its thinking. You can read that block to see exactly where the reasoning turned. For tasks with a single right answer, that visible pivot is often more useful than the answer itself, because it shows the model catching its own slip rather than burying it.

Because it builds on Llama 3.1 70B Instruct and keeps the standard Llama chat format, Reflection 70B drops into tools and pipelines that already support Llama. If you've deployed a Llama model before, the loading code and serving setup carry over with little change. That compatibility is why the model was easy for people to test in the first place.

The model targets tasks that reward accuracy over speed, like multi-step math and logic. It breaks reasoning into distinct steps and spends more tokens double-checking them. The tradeoff is latency. Answers take longer to arrive, and token costs climb, because reflection means writing more text before you get the final line.

The weights are publicly available, so you can run the model on your own hardware, fine-tune it, or study how reflection training changed its behavior. Open weights also mean you aren't locked into a single vendor's pricing or rate limits, which matters if you're testing the model at any scale.