Qwen3-4B-Thinking-2507-DES-Reasoning

Support our open-source dataset and model releases!

DES Reasoning: Qwen3-4B-Thinking-2507, gpt-oss-20b

DES Reasoning is an experimental specialist reasoning AI with custom output format; for general reasoning and chat, try Shining Valiant 3 for gpt-oss-20b!

DES Reasoning is a specialist reasoning assistant, performing situational analysis and reasoning to produce SimPy simulation scripts and strategies for analysis in response to user output.

• Finetuned on our DES dataset data generated with DeepSeek-V3.1!
• Multi-step analysis identifies the structure of the situation and the goal of simulation before proceeding to analysis and creating SimPy simulation code and analysis chat.
• DES Reasoning Format provides clear, readable Python code that is easy to read and modify; easy to use for running simulations, doing analysis, or further conversation with your assistant.
• Trained in a variety of subjects for flexible analysis: programming, science, business, economics, energy, finance, law, logistics, management, manufacturing, operations, supply chain and more!
• Small model sizes allow running on local desktop and mobile, plus super-fast server inference!

Prompting Guide

DES Reasoning uses the Qwen3-4B-Thinking-2507 prompt format to create SimPy Python scripts and chat analysis using DES Reasoning Format.

DES Reasoning is an experimental reasoning finetune:

• the assistant performs multi-step reasoning during the thinking phase, before producing the SimPy simulation code and expanded analysis chat for the user.
• describe the situation to be analyzed in order to prompt for the DES Reasoning Format; see the example script below for examples. Focus on making clear the goals you have and allow the DES Reasoning model to respond accordingly with analysis and simulation.
• this is an early experimental release: if used in a productive context, structural validation of outputs is strongly recommended.
• we recommend reasoning level high for all chats.

Example inference script to get started:

from transformers import pipeline
import torch

model_id = "sequelbox/Qwen3-4B-Thinking-2507-DES-Reasoning"

pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype="auto",
    device_map="auto",
)

prompt = "Compare two energy management strategies for a mobile CPU. The task is to process a batch of 10 million instructions. Strategy 1 (Race-to-Halt): Run the CPU at its maximum frequency (3 GHz, 1.2V, 5W power) to finish the task as quickly as possible, then enter a deep sleep state (0.05W). Strategy 2 (Pace): Run the CPU at a lower frequency (1.5 GHz, 0.95V, 1.5W power) that is just fast enough to meet a soft deadline of 100ms. The simulation must model the active time, sleep time, and state transition energy costs (100 microjoules to enter/exit sleep). The goal is to determine which strategy results in lower total energy consumption for the task."
#prompt = "A downtown bus terminal has 10 bays serving 20 different bus routes. The current system uses fixed bay assignments. This often leads to queues of buses waiting for their designated bay to become free, while other bays are empty. An alternative dynamic system is proposed where an arriving bus is allocated to the first available bay. Simulate both systems, considering bus arrival variability and passenger boarding times (which depend on the number of waiting passengers for that route). The objective is to compare the two systems based on bus turnaround time and passenger comfort."
#prompt = "An ERP system calculates an optimal production schedule daily based on an algorithm. However, an experienced human scheduler reviews and frequently modifies this schedule based on heuristics, such as prioritizing orders for historically difficult clients or building up just in case inventory ahead of a machine they don't trust. Model the production system under two scenarios: 1) The ERP schedule is followed perfectly. 2) The ERP schedule is subject to the scheduler's probabilistic overrides. Quantify the "cost of human intervention" by comparing the performance of the two systems on metrics like inventory cost, service level, and schedule stability."
#prompt = "A proposed system involves massive solar power stations in geosynchronous orbit that beam energy down to dedicated receiving stations (rectennas) on Earth via microwaves. Model the integration of a 2 GW baseload space-based solar power feed into a terrestrial grid. The simulation must account for the unique reliability characteristics: it is immune to weather and day/night cycles on Earth but is vulnerable to space debris impacts, solar flares, or beam misalignment, which could cause an instantaneous, total loss of power. The objective is to determine the required level of terrestrial spinning reserves to safely integrate this novel power source."

messages = [
    {"role": "user", "content": prompt},
]

outputs = pipe(
    messages,
    max_new_tokens=16000,
)
print(outputs[0]["generated_text"][-1])

DES Reasoning is one of our experimental reasoning releases; we've got more to come soon!

Do as you will.