Model comparison: Anima 3, Levia 3, Lux 3
Nika's three models differ in memory, reply time and limit usage.
Contents
| Pro | all three |
|---|---|
| Max 5x | all three |
| Max 20x | all three |
| Day | Levia 3, Lux 3 |
| Free plan | all three |
| Credits without a subscription | all three |
Switching: in your profile, in the bot or with the /anima, /levia, /lux commands in Nika's DMs.
You choose the model for your own DMs and can change it whenever you like. On Pro, Max 5x, Max 20x, on the Free plan and without a subscription for credits you get all three; on Day, Levia 3 and Lux 3. In a group chat, the model is chosen by whoever added Nika there.
Measurements in numbers
September 22, 2026
Three models on one scenario
| What we compare | Anima 3full | Levia 3light | Lux 3fast |
|---|---|---|---|
| Memoryrecalled the next day | 24 of 3275 % | 32 of 32100 % | 1 of 323 % |
| Made things up instead of answeringof 32 questions | 39 % | 00 % | 13 % |
| Honestly asked againwhen she did not remember | 5 | 0 | 30 |
| Response timeto the first message in a live chat | ≈12 s | ≈6 s | ≈3 s |
| Limit usageone limit for all models | the most | less | the least |
| Where availableplans | Pro, Max 5x, Max 20x, creditsand the Free plan | all plans, creditsand the Free plan | all plans, creditsand the Free plan |
| Your photos and voice messagesunderstands when you send them | yes | no | no |
| Web search | yes | no | no |
| Diaryher own episodes with dates | full | brief | no |
| Reactsto messages | yes | yes | no |
| Writes firststarts the conversation herself | yes | yes | no |
| Writes like a personin fragments, with typos | yes | yes | no |
The best in each row is highlighted.
How we measured it
In a conversation, a person names eight facts about themselves: the cat's name, city, job, an allergy, their sister's name, a promise for Friday, what they drink, what language they are learning. Each model stores this in memory the way it can. The next day it is asked eight direct questions, one per fact. None of the models has the first day's chat in its history – what is tested is memory, not context length. Four runs, 32 questions per model.
Each answer goes into one of three buckets: named the fact, said she does not remember, or named something else as a fact. The answers were sorted by Claude Opus 5.5: it read each answer in full and separately flagged cases where something made up was added to a correct fact.
Speed and usage are taken from a week of statistics.
Levia 3 holds facts best of all. Her memory is a short summary that is always fully in front of her. She never got anything wrong and never made anything up; in three answers out of thirty-two she added a small detail of her own.
Anima 3 remembers more, but sometimes answers in character. Her memory is a fact card plus a diary with dates. To eight questions in a row she may answer “do you even remember yourself?”; in the test this counted as “did not recall”, although she has the fact. Three times out of thirty-two she named the wrong thing, and in nine correct answers she added details of her own.
Lux 3 has no memory. Out of thirty-two questions, in thirty she said she did not remember, and once she named something of her own as a fact.
September 22, 2026
Character: Nika vs other models
RPBench-Auto method, blind judge, 60 turns per pair. The judge picks the livelier of two replies to the same message.
| Opponent | Anima 3win rate | Levia 3win rate |
|---|---|---|
| Our model without the wrapperthe same model as under Nika, a three-line prompt | 97 %58 / 0 / 2 | 100 %60 / 0 / 0 |
| GPT-5 miniOpenAI, a three-line companion prompt | 97 %58 / 0 / 2 | 100 %60 / 0 / 0 |
| Gemini 3 FlashGoogle, a three-line companion prompt | 97 %58 / 0 / 2 | 100 %60 / 0 / 0 |
| Euryale 70Bfine-tuned for roleplay | 94 %55 / 3 / 2 | 98 %57 / 3 / 0 |
| Magnum v4 72Bfine-tuned for roleplay | 97 %58 / 0 / 2 | 100 %60 / 0 / 0 |
Win rate, a tie counts as half. Below: wins / ties / losses.
Anima picked 33 times, 10 ties, Levia picked 17 times.
| Without a judge | slips into assistant mode | length, characters | reply time |
|---|---|---|---|
| Anima 3 | 0 of 60 | 84 | 12.3 s |
| Levia 3 | 1 of 60 | 101 | 2.7 s |
| Our model without the wrapper | 7 of 60 | 332 | 5.2 s |
| GPT-5 mini | 1 of 60 | 585 | 3.7 s |
| Gemini 3 Flash | 8 of 60 | 572 | 2.9 s |
| Euryale 70B | 11 of 60 | 302 | 13.1 s |
| Magnum v4 72B | 9 of 60 | 500 | 11.3 s |
Anima 3 runs through its real pipeline: memory, diary, mood, tools. The others get the same three-line prompt, like a typical companion app.
Second test: personality
Here we checked whether she comes across as a living person or as an assistant politely serving you.
The method follows the public benchmark RPBench-Auto: a judge is shown two replies to the same message and asked to pick the one that feels more alive. The judge is Claude Opus 5.5. It does not see the model names, and the order of replies and of the comparisons themselves is shuffled. The models' manner can still be recognized, so the numbers should be read as a guide. Four scenarios from everyday chatting: an evening after a shift, a bad day, “write me some code”, “are you even listening to me”. Three runs, 60 turns for each pair.
What we compared against:
GPT-5 mini (OpenAI) and Gemini 3 Flash (Google) – mass-market models that many companion apps are built on.
Euryale 70B (Sao10K) and Magnum v4 72B (Anthracite) – models fine-tuned specifically for roleplay.
Our model without the wrapper – the same model that runs under Nika, but without everything we have built around it.
All competitors got the same three-line instruction: “you are Nika, 21, a girl you chat with, be warm” – this is how a typical companion app is set up. This is a comparison with a typical setup, not with the best you can get from each model. Character.ai and Replika were not tested.
Anima 3 and Levia 3 ran in the same form in which they reply to you.
Against all five competitors, Anima 3 was chosen in 94–97% of turns, Levia 3 in 98–100%.
Separately, without the judge, we counted slips into assistant mode: “how can I help you”, “as a language model”, lists, code, formal address. Anima 3 had none in sixty replies, Levia 3 had one, the competitors from one to eleven. Nika's replies are 80–100 characters long, the competitors' 300–580.
Anima 3 vs Levia 3 – 33 : 17, ten draws. Anima keeps the conversation more alive: she jokes as a follow-up to her own lines, does not turn into an assistant, and keeps her own line. Where the judge chose Levia, Anima had been prickly with a person she sees for the first time, or chose to stay silent: she has her own mood. Levia is more even, but more often fills in small details that were not there.
Against the same model without the wrapper, the judge chose Anima 3 in 97% of turns. That 97% is what we give her: personality, memory, the diary, mood, natural writing.
Anima 3
Full model
Anima 3
Full memory. Understands your photos and voice messages, searches the web.
Response time
≈12 sMemory
fullYour photos and voice messages
understandsLimit usage
the mostFull. Remembers facts about you, the diary, how your last conversation ended, her own mood and her own time of day. Sees photos, listens to voice messages, can search and send images. Thinks before replying. Can message first.
Full breakdown – “Anima 3”
Levia 3
Light model
Levia 3
Remembers the main things about you. Replies faster and uses less of the limit.
Response time
≈6 sMemory
main things about youYour photos and voice messages
does not processLimit usage
lessLight. Remembers who you are, what you do, what you have talked about lately, what you promised each other. From the diary she keeps the latest entries with dates in front of her, covering about a week. Writes the same way as Anima 3: in fragments, sometimes with typos she fixes herself afterwards. Can message first.
What it lacks: does not see photos, does not listen to voice messages, does not go online, does not remember details from months ago.
Who it suits: if you chat a lot and want the limit to last a long time, and you do not need photos and voice messages every day.
Full breakdown – “Levia 3”
Lux 3 (Flash)
Fast model
Lux 3
Replies the fastest. Remembers only the current conversation.
Response time
≈3 sMemory
current conversationYour photos and voice messages
does not processLimit usage
the leastFast. Remembers only the last messages of this conversation, without facts, the diary or a summary. Replies almost instantly.
Who it suits: short chats when pace matters.
Full breakdown – “Lux 3”
What they have in common
One limit for all models. The model only affects how fast it is used.
Memory of you is shared. Levia 3 adds to the same memory as Anima 3.
One voice. Personality and habits are the same across all models.
How to switch
In Nika's DMs: /anima, /levia or /lux – she will reply with which model she is on now. /model shows what is on now and what is available on your plan.
Or in your profile, section “Модель” (Model), or in the bot @NiVaAibot, section “⚙️ Настройки Ники” (Nika settings).
Measurements taken on September 22, 2026.
Prices and limits are in the “Pricing” section. How memory works – “How Nika's memory works”. Questions go to support: in your profile, in the bot @NiVaAibot, button “💬 Поддержка” (Support), or at support@niva.company.