beezu/Magistry-24B-v1.1-mlx-mxfp4
This model beezu/Magistry-24B-v1.1-mlx-mxfp4 was converted to MLX format from sophosympatheia/Magistry-24B-v1.1 using mlx-lm version 0.31.2.
Original Model Card
Magistry-24B-v1.1
A Royal Merge · 24B · Apache 2.0
Please see my model card for sophosympatheia/Magistry-24B-v1.0 for background information on this line of models.
Magistry-24B-v1.1 was the result of my attempts to improve on the coherency and handling of v1.0 without compromising what makes it fun. After performing some dark rites in my laboratory, I think I managed to produce a worthy successor. This version of Magistry feels like a different beast. It is hard to say whether it is truly "smarter" than v1.0, but it at least writes like it wants to be seen that way. Is it? Test it for yourself and let me know. It didn't lose any of its pizzazz, that's for sure. If anything, I think Devstral somehow made this version even more creative and entertaining.
Known Issues
This model is fun, but it will struggle with details of logical/physical continuity at times — which is probably inescapable for a 24B model. Rerolling the output might fix it, or you might have to help it out by providing explicit instructions or 'reminders' so it doesn't get confused. It should at least outperform v1.0 in this area.
Thinking can also help this model with continuity and planning, but don't expect miracles. If you force its reply to start with <think> and give it some basic instructions, it can totally do it. I find <think></think> tags work better than Mistral's [THINK][/THINK] tags.
<think>
First draft your thinking process (inner monologue) until you arrive at a response.
Your thinking process must follow the template below: Your thoughts or/and draft, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate the response. Responses must develop logically from past context.
THINKING:
Sampler Tips
You can use the master import JSON in this repo (Magistry_SillyTavern_Master_Import.json) to deploy the conservative sampler settings below, which are likely to be compatible with more backend/frontend combos. I recommend using these values as a starting point for your own experiments. It's not like the model falls apart if you deviate from these settings, but they should be a reliable starting point for most creative tasks.
Conservative Settings
Run these settings as a starting point if you care more about accuracy than flair. If you can't use Top-N σ, you can safely drop it for ubiquitous samplers like Top-P and Top-K, and then tune those settings plus Temperature and Min-P to get virtually the same effect.
Temp
0.7
Min-P
0.05
Top-N σ
0.75
DRY Mult.
0.8
DRY Base
1.8
Wild Settings
Run these settings as a starting point if you care more about flair and fun than accuracy. You will have to reroll more often and deal with shenanigans, but they'll be entertaining most of the time. These Adaptive-P settings differ from what is usually recommended, but trust me on this: it's fun with this model and works fine for simple scenarios and conversations. The low Adaptive-P Decay setting frees the model to "fishtail" between wild token choices and conservative token choices in a way that mostly keeps things on track while unleashing creativity. If you need stability for complex scenarios, then run either the Conservative Settings above or the Balanced Settings below.
Temp
0.9
Min-P
0.05
Adaptive-P Target
0.35
Adaptive-P Decay
0.45
Balanced Settings
Run these settings as a starting point if you want a reasonably stable balance of flair and accuracy. Adaptive-P seems to work wonders with this model, and you can tune the parameters to get more or less predictable results. (See the creator's GitHub docs for more information on Adaptive-P parameters.) If you need more stability, lower Temperature and increase Adaptive-P Target. If you want more creativity, lower Adaptive-P Target. Tune Adaptive-P Decay last since the default 0.9 is recommended and stable.
Temp
1.0
Min-P
0.05
Adaptive-P Target
0.6
Adaptive-P Decay
0.9
Prompting Tips
You can download the Magistry_SillyTavern_Master_Import.json file from this repo and import it directly into SillyTavern to get system prompt, chat template, and sampler settings all in one go.
Donations
If you feel like saying thanks with a donation, I'm on Ko-Fi
Quantizations
License
Apache 2.0, inherited down from Magistral.
Merge Details
This is a merge of pre-trained language models created using mergekit.
Merge Method
This model was merged using the DELLA merge method, using Darkhn/Magistral-2509-24B-Text-Only as a base.
Models Merged
The following models were included in the merge:
Configuration YAML
models:
- model: Darkhn/Magistral-2509-24B-Text-Only
# base model — no parameters needed
- model: Casual-Autopsy/Maginum-Cydoms-24B
parameters:
weight:
- filter: "self_attn"
value: 0.55 # pull attn closer to base — preserves magistral's reasoning
- filter: "mlp"
value: 0.8 # keep mlp weight — creative energy
- value: 0.8 # fallback for embed/norm/lm_head
density: 0.9
epsilon: 0.099
- model: DarkArtsForge/Magistaroth-24B-v1
parameters:
weight:
- filter: "self_attn"
value: 0.55 # pull attn closer to base — preserves magistral's reasoning
- filter: "mlp"
value: 0.8 # keep mlp weight — creative energy
- value: 0.8 # fallback for embed/norm/lm_head
density: 0.9
epsilon: 0.099
- model: huihui-ai/Huihui-Devstral-Small-2-24B-Instruct-2512-abliterated # Note: this cannot be merged directly without significant pre-processing
parameters:
weight: 0.3
density: 0.9
epsilon: 0.099
merge_method: della
base_model: /home/llm/mergequant/models/BASE/Darkhn_Magistral-2509-24B-Text-Only
parameters:
lambda: 1.0
normalize: false
tokenizer:
source: union
chat_template: auto
dtype: float32
out_dtype: bfloat16
