license: apache-2.0
library_name: transformers
base_model:
- mistralai/Mistral-Small-3.2-24B-Instruct-2506
- mistralai/Magistral-Small-2509
tags:
- merge
- mergekit
- slerp
- mistral
- reasoning
- 24b
language:
- en
- fr
- de
- es
- it
- pt
- zh
- ja
- ko
- ar
Taipei 2
A 50/50 SLERP merge of Mistral-Small-3.2-24B-Instruct-2506 and Magistral-Small-2509, both 24B Mistral-3 architecture models sharing the same base. This has resulted in our best model, Taipei 3.1
The goal: combine the conversational polish, tool-calling reliability, and low-latency response style of Mistral Small 3.2 with the explicit reasoning capability (SFT + RL on Magistral Medium traces) of Magistral Small 1.2. The merged model retains the [THINK]/[/THINK] reasoning tokens from Magistral via tokenizer_source: union, so it can operate in either fast-response or deep-reasoning mode depending on system prompt.
Use
Works with vLLM, transformers, and llama.cpp (after GGUF conversion). Use Magistral's system prompt format to enable reasoning traces; use a standard Mistral system prompt for fast chat.
Tokenizer
This repo ships Mistral's canonical tekken.json rather than a serialized HF tokenizer.json. transformers' AutoTokenizer.from_pretrained auto-converts it on load. For best fidelity in production, use mistral-common or vLLM, which read tekken directly. The [THINK] / [/THINK] reasoning tokens are preserved (ranks 34 / 35).
Merge config
merge_method: slerp
base_model: mistralai/Mistral-Small-3.2-24B-Instruct-2506
slices:
- sources:
- model: mistralai/Mistral-Small-3.2-24B-Instruct-2506
layer_range: [0, 40]
- model: mistralai/Magistral-Small-2509
layer_range: [0, 40]
parameters:
t:
- filter: self_attn
value: [0, 0.5, 0.3, 0.7, 1]
- filter: mlp
value: [1, 0.5, 0.7, 0.3, 0]
- value: 0.5
embed_slerp: true
dtype: bfloat16
tokenizer_source: union
Part of the Tripplet Taipei model series.