Llama 3 8B Lunaris: What a Four-Model Merge Is Good For

Lunaris is a small model with a narrow job: casual chat and roleplay at a price low enough to stop thinking about. It is an 8B model based on Llama 3, made by Sao10K, and it is not trained so much as assembled. That difference explains both why it exists and where it breaks.

If you came here comparing it with frontier models, the comparison is not useful. The useful questions are whether an 8B model with an 8K window can carry your workload, and whether a merge behaves predictably enough to put in front of users.

A merge, not a fine-tune

The Hugging Face card describes Lunaris as a TIES merge on top of Meta's Llama 3 8B Instruct. Instead of training on new data, a merge combines the weight changes of several existing fine-tunes into one model. TIES is a method for doing that without the changes cancelling each other out: it keeps only the largest changes from each source, resolves sign conflicts between them, and then adds the survivors together.

Lunaris combines four sources, each with a published density and weight:

• an earlier roleplay model by the same author, Stheno v3.2, given the largest weight;
• a megamerge the author includes because it helps with general knowledge;
• two further community models tuned for roleplay and storytelling.

The author's stated reasoning is that the other models balance out Stheno v3.2's creativity while improving its logic, and that merging models built on different data works better than training on all of it at once. The author also says the merge values came from long-running personal experimentation. That is honest, and it is also the caveat: a merge is tuned by feel, and there is no training run whose loss curve tells you what it learned.

8K context is the constraint that matters

OpenRouter lists sao10k/l3-lunaris-8b with an 8,192-token context window, which matches the original Llama 3 8B base. That is the number to plan around.

For roleplay, 8K fills quickly. A character card, a system prompt and a lorebook entry can take a large share of it before the first message. After that, every turn pushes older turns out. The model does not warn you when this happens; it just starts forgetting names and earlier events. If your app is built on long sessions, you will need history summarisation, and it has to be good, because the model has very little room to recover lost context from.

Newer 8B bases have far larger windows. If context length is the deciding factor, an 8B model on a newer base is the more sensible starting point than an older merge.

The price changes what you use it for

At the time of writing OpenRouter listed Lunaris at $0.04 per million input tokens and $0.05 per million output. At that rate the model cost of a casual chat product rounds to nothing, and hosting, moderation and support become the real bill.

That makes Lunaris a candidate for jobs where volume is high and stakes are low:

• background characters in a game, where many short lines are generated and none is critical;
• a first-pass draft that a larger model or a human then edits;
• the cheap tier in a routed system, where a bigger model takes over when the conversation needs it.

That last pattern is the strongest fit. Model routing and fallbacks covers how to send easy turns to a cheap model and escalate the rest, and an 8B creative model is a natural bottom rung.

Lunaris against the 70B Euryale models

The same author's 70B line, the Llama 3.1 Euryale v2.2 and its Llama 3.3 successor, cost roughly fifteen to twenty times more per token as listed on OpenRouter at the time of writing, and offer 131K context windows against Lunaris's 8K.

• Pick Lunaris for short exchanges, high volume, tight budgets, or self-hosting on a single consumer GPU, where an 8B model is practical and a 70B model is not.
• Pick a 70B Euryale when sessions are long, when characters must stay consistent across many turns, or when a scene involves several characters and spatial detail. Small models tend to lose track of who is where sooner.

Licence also differs. The Lunaris repository lists the Llama 3 licence, which is Meta's community licence rather than a non-commercial Creative Commons one. Read it against your use, as with any Llama derivative.

Settings and a test plan

The card recommends temperature 1.4 with min_p 0.1 and the Llama 3 Instruct template. Temperature 1.4 is very high; it only works because min_p prunes unlikely tokens relative to the most likely one. Without min_p, start closer to 1.0. Small vs large models covers where a model this size is enough and where it runs out.

To decide whether it is good enough:

1. Take ten real conversations from your product, or write ten that resemble them.
2. Run each to the point where the context window fills, then a few turns past it.
3. Note the turn at which the model first contradicts an earlier fact. That is your practical session length.
4. Compare cost per conversation against a 70B model at the same quality bar, including regenerations.

If the practical session length covers most of your conversations, Lunaris is a bargain. If it does not, the low price will not make up for it.