aimeri/spoomplesmaxx-thrasher-24B

SpoomplesMaxx-Thrasher-24B

"Thrash Metal"

▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ ░░░░░ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░██ ▒▓██▓▒  ░    ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░      ▓░▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░▓▓▒█▓▓▓▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ ░░░░░█▓░▓█ ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░▒███▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░░▒░░░░▓ ▓ ▓█▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░▓░ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▒░░ █▓ █▓▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ ░░░░░░░▒░░▒▓  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▒░░░░  █ ▓▓▓ ▓▓█▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░ ░░░░░   ▒░▓▓░ █▓▓▒▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░    ▓▓ ▓    ▓  ▓▓▓ ▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓         █▓▓ █▓▓▓ ▓▓█▓█ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒  █         ▓ ░▓▓▓▓ ▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒   ▓  ▓▒▓▓ ▒▓▓▓▓▓▓▓█ ▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓░ ▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░ ▒       ▒ ▓ ▓▓ ▓▓ ▓▓▓ ▓▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓   ░ ░  ░░▓▓▓▓ ▓▓░▓▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  ░    ░░░▓▒█▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓      ▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓     ▓▓▓░░░░▓▓ ▓▓▓▓ ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓   ▓▓▓░   ▓▓▓▓  ▓  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓   ▓▓█▓▓▓▓▓ ▓▓▓▓▓▓▓▓ ▒▓▓▓▓▓▓▓▓▓▓▓▓ ▓▓▓▓      ░▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓    ▓▒▓▓▓▓▓▓▓▓ ▓▓▓▓▓▓▓▓▓ ▓▓▓▓▓▓▓▓░     ░         ▒▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓  ▓▓▓▓▓▓▓▓▓▓      ▓▓▓▓▓▓▓▓▓▓ ▓▓▓▓▓▓▓▓▓ ░ ▓        ▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓▓ ▒▓ ▒ ▓▓▓▓▓▓▓▓▓▓▓ ░▓▓▓▓▓▓▓  ▓   ▒▓▓▓▓▓▓    ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓  ▓ ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓  ▓▓        ▒ ▓▓▓▓▓▓▓      ▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓   ▓▓▓▓   ▓ ░▓▓▓▓▓▓▓▓▓▓▓▓▓▓        ▓▓▓▓ ▓▓▓▓▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓  ▓▓▓     ░▓▓▓▓▓▓▓▓▓▓▓▓▓      ▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓░   ▓▓ ▓▓ ▓  ░░▓▓▓▓▓▓▓▓▓▓▓      ▓▓ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  ▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓            ▓ ░ ▓▓▓▓▓▓▓▓▓     ▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓   ▓ ░░▓▓▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓   ▒ ░ ▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓ ░   ░▓  ▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓     ░▓▓▓       ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓░ ▓  ▒▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓    ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓     ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
                            

"Thousand-Songed" — Toxostoma rufum, the brown thrasher: holder of the largest documented song repertoire of any North American bird, over a thousand song types. The mockingbird repeats a phrase three times; the thrasher sings each one twice and moves on. Second of the mimids, the family that follows the corvids.

A model that is 100% about roleplay, trained mostly on things that are not roleplay. Measured across the strongest open RP lineage I know (Dans-PersonalityEngine), roughly 590K of its rows are task/reasoning/assistant/world-knowledge data against ~150K of actual roleplay. RP is the product; RP is not the corpus. The RP data teaches the register. Everything else teaches the mind behind it.

Built on Mistral-Small-3.1-24B-Base with the vision tower removed and the chat interface rebuilt from scratch — see "The token surgery" below, because if you have ever bounced off a Mistral model's template, that section is for you.

Who this is for

The accessible mimid. mockingbird at 36B asks a lot of your VRAM; thrasher at 24B (23.6B after the vision strip) is the same recipe in a size that quantizes onto a single 24GB card. If you can run any of the popular 22–24B RP models, you can run this one.

Prompt format

ChatML. On a Mistral base. Yes, really — and not by resizing anything:

FORM     <|im_start|>system\n{card}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant\n{reply}<|im_end|>\n
STOPS    <|im_end|>   (id 21; eos)
BOS      <s> (id 1) — added automatically by the tokenizer, not the template
ROLES    system / user / assistant / tool
EXAMPLE
<|im_start|>system
You are Bram Hollis, keeper of the Wayward Lantern...<|im_end|>
<|im_start|>user
*I push the door open, dripping wet* Got room for one more?<|im_end|>
<|im_start|>assistant

The template ships embedded (chat_template.jinja + tokenizer_config.json), so vLLM, llama.cpp, MLX, and every frontend that speaks ChatML — which is to say, every RP frontend — picks it up without ceremony.

No thinking. Ever.

thrasher never emits reasoning traces and was never trained on them.

The token surgery

Moving a Mistral base to ChatML was my first barrier the first time I ever tried to use a Mistral model for fine-tuning, so here is the whole recipe, with receipts.

1. There is room in the vocabulary. Mistral's Tekken tokenizer reserves ids 0–999 as a control block; only 0–19 are named. Ids 20–999 are unused <SPECIAL_n> placeholders. So ChatML needs no vocab resize and no embedding-matrix growth: rename <SPECIAL_20> → <|im_start|> and <SPECIAL_21> → <|im_end|> in tokenizer.json (both added_tokens and the vocab), tokenizer_config.json, and special_tokens_map.json. Set eos to <|im_end|>. Done — single-token ids 20 and 21.

2. The claimed rows are dead, and dead eos rows are fatal. The base was pretrained with those placeholders never appearing in data: their embedding rows are exactly 0.0 (measured), their lm_head rows random-scale noise. A model whose eos row is dead never learns to stop — unless you train very long (Hermes cold-claimed on 60B tokens; this SFT is ~1B) or you initialize sensibly. thrasher grafts: <|im_start|> rows ← <s>, <|im_end|> rows ← </s>, embedding and lm_head. By step 100 of SFT, stop-rate at temperature 0.7 was already 24/24; it stayed 100% at every checkpoint probed.

3. The published tokenizer has a broken pre-tokenizer regex. The known Mistral conversion bug (transformers warns and offers fix_mistral_regex=True — but that fix is in-memory only, and axolotl, llama.cpp, and MLX read tokenizer.json directly). The true Tekken pattern — case-aware word splitting, single-digit number splits — is baked into the shipped file.

4. The vision tower is gone. 222 tensors of Pixtral removed, the language_model. prefix stripped, published as a plain untied MistralForCausalLM. Nothing multimodal remains.

The full prep script (prep_base.py) is in this repo, and the prepped base is its own artifact if you want to start from it: Mistral-Small-3.1-24B-Base-thrasher.

Tool calling

The corpus includes the full Toolmaxx family (58,095 conversations), rendered with tool responses as a plain tool role turn:

<|im_start|>tool\n{tool output}<|im_end|>

Corpus-taught, conversational tool competence — not a structured calling API.

If you need strict function calling, put a schema in the card and validate what comes back.

Key details

BASE       mistralai/Mistral-Small-3.1-24B-Base-2503  (Apache 2.0, vision stripped)
PARAMS     23.6B dense · 40 layers · GQA 8 KV heads · head_dim 128 · hidden 5120
VOCAB      131,072 · ChatML on claimed Tekken slots 20/21 · zero added tokens
CTX        trained at 24,576 packed · base RoPE (theta 1e9) to 131K
CORPUS     667,332 conversations · ~1.3B supervised chars · 43% RP share
LANGUAGE   English (non-English filtered at ingest; base priors remain)

Training

Full-parameter SFT, Axolotl, 8×H200. One stage. The exact config generator ships in this repo (thrasher_sft_cfg.py):

STEPS     902 (2 epochs) · this release = step 600
SEQ       24,576 · sample packing (block-masked; packing does not shrink context)
BATCH     64 global (micro 1 × accum 8 × 8 GPUs)
OPT       AdamW · lr 8e-6 cosine · 3% warmup · wd 0.01 · bf16
STACK     FSDP2 full-shard · activation checkpointing · Cut Cross Entropy
HEALTH    grad_norm 1.2–1.8 the whole run · zero spikes · memory flat

On context: 24,576 is the longest single training conversation (longer ones were split at turn boundaries with the card re-carried). The base's 131K RoPE survives SFT untouched; the best-trained RP region is the first ~24K, degrading gracefully beyond.

The corpus is mockingbird's, verbatim — the PersonalityEngine V1.3.0 public list plus my own carded-RP, think-stripped-RP, and anti-repetition lanes, same cleaning receipts. Only the template changed. See the mockingbird card for the full corpus story.

How the checkpoint was chosen — and how you can check

Loss did not pick this model. Checkpoints went through two instruments, both published in this repo's eval/: a seeded multi-turn loop/stall battery (six gates, six repeats per episode — single-run numbers on it are noise, and the trend proves it), and blind-judged episodes on four real character cards.

step        100   200   300   450   600   750   900
battery      17     8    13    13    18    10    16   (of 24)
stop-rate   100%  100%  100%  100%  100%  100%  100%

The shape is the story: a mid-run dip while style reorganizes under high LR, a peak mid-way through epoch 2, then regression in the deep anneal. Behavioral peak ≠ end of training. Step 600 shipped — highest battery pass rate and the most disciplined judged transcripts (compressed, declarative, no user-impersonation).

The anti-repetition anneal experiments — published, not spun. After training, we annealed checkpoints 600 and 900 on a 7,281-conversation anti-repetition dataset (recipe in that repo), across four blend/base variants plus two task-arithmetic merges. In our measurements, one variant reached zero repeat failures on the battery while regressing on format gates; the merges regressed stop-token reliability; none beat the un-annealed step 600 overall, so step 600 shipped as-is. The raw battery JSONs for every variant are in eval/ — read them and draw your own conclusions rather than taking ours.

Sampling

The shipped generation_config.json is the swept optimum (7 arms × 5 seeded battery runs each):

temperature 1.0 · min_p 0.05 · top_p off

thrasher is the anti-mockingbird in its sampler behavior, which is why we sweep per model instead of inheriting: min_p won here (mockingbird's sweep found min_p looped MORE); and a mild repetition_penalty 1.05 — catastrophic on mockingbird — is a legitimate opt-in on thrasher: in our sweep it eliminated the verbatim-loop tail entirely (max cross-turn Jaccard 0.50 across 20 episodes) at the cost of a rare unfinished turn. If loops bother you more than an occasional run-on, add it. Plain top_p 0.9 at temp 1.0 was the worst repetition arm on this model — don't ship RP muscle memory, sweep.

Known limitation: the loop tail.

Under sustained low-information multi-turn pressure the model can fall into near-verbatim self-repetition — roughly 1–2 episodes in 24 on our battery at shipped settings. The corpus's anti-repetition lane suppresses it; it is not eliminated. The repetition_penalty 1.05 opt-in above removed it entirely in our measurements.

Give it a proper card and it will give you a proper character: the model was fed real character cards (median ~3K chars, p90 ~8.5K) as system messages.

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "aimeri/spoomplesmaxx-thrasher-24B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
messages = [
  {"role": "system", "content": "You are Bram Hollis, keeper of the Wayward Lantern... Third person, *asterisk action beats*."},
  {"role": "user", "content": "*I push the door open, dripping wet* Got room for one more tonight?"},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=400, do_sample=True)  # sampler ships in generation_config
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Quants: GGUF static (Q3/Q4/Q5_K_M) · GGUF imatrix (IQ3_XXS–Q4_K_M, own-corpus calibration, imatrix.dat included) · MLX 4-bit / 6-bit. Prefer the imatrix quants at 3–4 bit.

The thrasher knows a thousand songs. It only needs the one you hand it.

thrasher is a roleplay and creative-writing model for adults. It stays in character by design — its corpus was scrubbed of mid-scene refusals — so bring your own moderation where your deployment needs it. Not an assistant, not an oracle, not for anything safety-critical.

mimids 02 · trained 2026-08 · checkpoints at thrasher-v1-ckpts · eval instruments, prep scripts, and training config in this repo · Apache 2.0