Latest Anima Derivative Checkpoints as of October 2026, from 2B Updates to 2.9B and 3.8B Merges
Contents
In June I wrote a list of 20+ Anima derivative checkpoints, and many of the versions in it have since been updated.
I have tested several of these in separate posts: character LoRAs on six derivatives, plain Anima-2.9B, Anima-3.8B, and official v1.1. This post looks up which version of each model is currently being distributed.
On October 21, 2026, I pulled the Civitai pages of the derivatives covered in earlier posts, along with the top 40 Anima checkpoints by monthly downloads, again through the Civitai API.
The query was baseModels=Anima&types=Checkpoint&sort=Most Downloaded&period=Month&limit=40&nsfw=true. According to the Civitai API reference, period=Month sets the time window used for sorting, baseModels=Anima filters by the registered base model, and nsfw=true includes adult models.
These 40 are not every Anima derivative, only the top 40 results for this query. The stats.downloadCount value the API returns is the all-time total for the model page, not the number behind the monthly ranking.
Dates are Civitai publish dates (UTC). The “Author’s notes” columns come from the model pages. I did not verify them with my own outputs.
I generated images from the current versions of the derivatives in the June list under identical settings in separate posts: seven 2B models and four 2.9B models.
A few terms that come up throughout:
| Term | Meaning in this post |
|---|---|
| 2B / 2.9B / 3.8B | Size labels on the model’s DiT (the image-generation network). Anima-Base has 28 blocks, Anima-2.9B has 40, and Anima Delta Mix 3.8B has 52 |
| Block layout and LoRA | A LoRA’s weights are tied to block numbers. If blocks are inserted, the numbering shifts, and a LoRA trained on Anima-Base has to be remapped |
| bf16 / fp16 / int8 | Weight precision. int8 files are smaller than 16-bit ones |
| Text encoder / VAE | The text encoder turns the prompt into input for the model (Anima’s standard one is Qwen3-0.6B). The VAE turns the model’s output into an image. Some pages bundle them into the checkpoint, others distribute them separately |
Current official Anima versions
As of October 21, the newest release on the official Civitai page was still Turbo v1.1 from August 24.
| Version | Published | Author’s notes |
|---|---|---|
| base-v1.0 | May 14 | Base for LoRA training. Still the current Base |
| aesthetic-v1.1 | July 13 | Current Aesthetic, with part of the model retrained to adjust the art style |
| turbo-v1.1 | August 24 | Current Turbo, with most of the distillation and style-tuning parts retrained |
In the Turbo v1.1 vs. Aesthetic v1.1 comparison I also tried Base-trained character LoRAs. The comparison images and settings, including the older versions, are in that post.
Current versions of derivatives from the June list (2B)
These are the derivatives from the June list that got a new 2B release in September or later, showing only the latest version as of October 21.
If a version keeps the same 28-block layout as Anima-Base, character LoRAs made for 28 blocks load without renumbering. Whether the derivative reproduces the same character is a separate question. In my June test of six derivatives, the trigger word alone did not give stable results.
| Model | Latest version | Published | Author’s notes |
|---|---|---|---|
| Hikari Anima | v1.2.0 | October 16 | Fewer broken compositions. The author writes that depth and detail around the eyes “feel improved” and are “easy to notice” |
| AnimaIka | v4.9 | October 7 | Less instability on seeds that produce multiple characters. Slightly flatter look |
| Hotaru Blend Anima | v6.0 | September 29 | Rebuilt from scratch, with a detailed, slightly cute, vivid style. Turbo LoRA baked in. An int8 convrot version is also available |
| Anima Cat Tower | v1.1 | September 29 | More stable style with natural-language prompts. Captions were redone with Qwen 3.8/3.6 or Gemma 4, and every one was checked by hand |
| CapAnima | Turbo 1.2 | September 25 | Style fixes. The Aesthetic version with a Turbo LoRA applied |
| RDBT | v2.3 16s | September 19 | No version notes. Other v2.3 variants such as base remain on the same page. v3.0 base sma came out on October 22 |
| AnimaYume | v1.5 (base) | September 6 | No notes |
Hotaru Blend v6.0 recommends Euler A or ER SDE, the Normal scheduler, 20–35 steps, and CFG 2.5–3.5 (8–16 steps and CFG 1.0–1.5 with Turbo).
The outputs from these seven models with a Base-trained character LoRA applied are in the seven 2B models post.
The move to Anima-2.9B
Since June, more derivatives have been labeled 2.9B.
According to the Anima-2.9B model card and the manifest that records where the layers were inserted, 12 blocks are inserted between Anima-Base’s 28 blocks, for 40 blocks in total.
In preview v1, only the 12 new blocks were trained, and the original 28 blocks were frozen.
Among the top 40 by monthly downloads, 18 had “2.9B” in the model name or the latest version name, and 2 had “3.8B”.
One of those 18 has both 2B and 2.9B in its model name, but its latest version is 2B. I treat the remaining 17 as current 2.9B releases. This classification is based on names and author notes, not on inspecting the contents of every checkpoint.
Derivatives from the June list also have 2.9B versions now: Nova Anime AM, CottonAnima, JANIMA, and AMIXA. JANIMA and AMIXA are not in this top 40, so I checked them individually from the June list.
The post where I tested plain 2.9B covers generation times on an M1 Max and how to remap Base-trained LoRAs to the 40-block layout.
Current versions of derivatives from the June list (2.9B)
| Model | Latest version | Published | Author’s notes |
|---|---|---|---|
| CottonAnima | 2.9B_v1 | October 15 | Fixed-style merge. Distributed in bf16 and int8. The author recommends ER SDE, 24–30 steps, CFG 5 |
| AMIXA2.9B | 2.9B_V01 | October 12 | A 5.44GB fp16 model-only file and a 7.12GB bf16 file that includes the VAE and text encoder. Recommends DPM++ 2M + Beta, 23 steps, CFG 4 |
| Nova Anime AM | v6.0:2.9B | October 9 | Built from Anima Base v1.0 + Aesthetic v1.1 + 2.9B. Claims better handling of prompts and context. Recommends Euler a + Normal, 20–30 steps, CFG 4–6 |
| JANIMA | v1.0 2.9B | September 8 | The author says the new 2.9B layers were merged in. Distributes bf16 and int8 weights, with the text encoder and VAE as separate files |
The settings in the table are the authors’ recommendations. The results of generating with all four under identical settings are in the four 2.9B models post.
After the V1 I tested in June, AMIXA’s model page was renamed to AMIXA2.9B. JANIMA’s 2.9B release also uses different weights from the 2B version I tested in June.
The CottonAnima author tells users to install ComfyUI-Anima-2.9B.
However, the Anima-2.9B author’s instructions say ComfyUI v0.33.1 and later load the 40 blocks out of the box. You only need the compatibility patch on ComfyUI versions older than v0.33.1.
2.9B releases in the top 40 by monthly downloads
Of the top 40 I pulled, 17 were classified as current 2.9B releases. Nova Anime AM and CottonAnima are in the table above, so the remaining 15 are listed here with their latest versions only. Unholy Desire Mix has both 2B and 2.9B in its model name, but its latest version is 2B, so it is not in this table.
| Model | Latest version as of October 21 | Published (UTC) |
|---|---|---|
| MiaoMiao Harem | 2.9B_BETA1.1 | September 26 |
| MiaoMiao RealSkin | Anima2.9B 1.0 | October 19 |
| Fn-Moment Anima-Turbo | V5.0-2.9B-no turbo | October 16 |
| Nova Furry AM | v4.0:2.9B | September 21 |
| Nova Orange AM | v5.0:2.9B | October 16 |
| Nova 3DCG AM | v3.0:2.9B | September 29 |
| PVC Style Model | Anima 2.9B 1.0 | October 13 |
| One obsession Branch | Msture anima_2.9B木珠 | October 18 |
| Nova Reality AM | v2.0:2.9B | September 23 |
| Anima Delta Mix 2.9B | DeltaMix v7.0 Flat | October 19 |
| Nova EXAnime AM | v2.0:2.9B | September 25 |
| UnNamedA | v2.0-t-2.9B | October 7 |
| Nova Mature AM | v2.0:2.9B | September 22 |
| Nova Cartoon AM | v2.0:2.9B | September 30 |
| uwumerge Anima Edition | V9 [Anima-2.9B+Aesthetic] | October 1 |
The seven Nova models are a series by the same author as Nova Anime AM, and all of them moved to 2.9B-labeled versions in September. For the other 20 models, the model name and latest version name are not enough to tell 2B from 2.9B, so I did not classify their size. You can check each model in the API query results linked at the top.
The two 3.8B models are not necessarily built the same way
The two 3.8B-labeled models in the top 40 are boleromix v0.14 and Anima Delta Mix 3.8B v6.0.
The Delta Mix author states that it uses a 52-block DiT but only Anima’s standard Qwen3-0.6B text encoder. The page warns against adding Qwen3.5-4B or an extension adapter, and points users with older LoRAs to a dedicated patch.
In contrast, the Anima-3.8B model card and my August post testing the preview use a setup that adds Qwen3.5-4B. Even though both say “3.8B,” you cannot carry that post’s workflow or LoRA setup straight over to Delta Mix. The boleromix model page does not specify a text encoder, so I could not tell how it is built.