Thinklite-27B

Thinklite-27B

A Qwen3.8-27B variant that reasons with far fewer tokens, so final answers arrive about twice as fast at high effort. For people who run reasoning models and want shorter waits.

Made by Sparset on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Thinklite-27B is a variant of Qwen3.8-27B built to spend fewer tokens on reasoning. By the card's own averages over six benchmarks, it cuts reasoning by 47% at the highest effort setting and still scores 89.8 against the base model's 90.2. It takes images and text and returns text.

Who it suits: It suits developers and teams who run a reasoning model and care about how long users wait for the final answer.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a reasoning model locally with shorter waits | A coding helper that answers sooner
Agent building | Agent builders | Cut the cost of long reasoning in multi-step agents | A tool-using agent that finishes each step faster
AI research | Researchers | Compare efficient-reasoning methods | Test it against ThinkingCap and Swift variants of the same base
Operations | IT teams | Host one model for staff questions | An internal assistant with shorter response times

If you want the base model's longest reasoning on every question, you can skip it.

Where to use it:

• Python with the transformers library
• llama.cpp and Ollama, through the separate GGUF release
• Windows, macOS or Linux machines with enough memory for a 27B model

Use case | Who | What they get back
Hard questions | Developers | A final answer after fewer reasoning tokens
Quick replies | App teams | Shorter waits at the low effort setting
Image and text prompts | Researchers | Text answers about pictures and documents

How to set it up: You need these first:

• Python and pip
• A GPU setup with room for about 55.6 GB of BF16 weights, or about 13.9 GB at a 4-bit estimate

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load the model with transformers as the model card describes.
4. For llama.cpp or Ollama, use the separate GGUF release instead, which is a 16.5 GB Q4_K_M file.

pip install -U huggingface_hub
hf download Sparset/Thinklite-27B

How to use it: Ask your questions as you would for Qwen3.8-27B, and choose the reasoning effort. The savings apply at both settings the card reports, xhigh and low. The model card holds the exact way to run it.

Tool | What it does
reasoning_effort xhigh | Cuts reasoning by 47% on average and makes the final answer about 2x quicker
reasoning_effort low | Cuts reasoning by 28% on average, with a 1.3x faster final answer
TTFO | The time until the complete final answer arrives, used to measure speed

Example: You ask a hard maths question at the xhigh setting. By the card's averages, the model uses about half the reasoning of the base model, so the final answer comes back in roughly half the wait. Single questions will differ.

For newcomers and experts: A reasoning model thinks in many tokens before it answers, and that takes time. This one aims to think shorter without losing accuracy. A first small thing to try is the same question at the low and the xhigh setting.

The reasoning saving counts average reasoning tokens per answer, compared with Qwen3.8-27B at the same effort setting, across six benchmarks. It also reports the speed-up in time to final output, and it compares with ThinkingCap-Qwen3.8-27B and Swift-1.5-Qwen3.8-27b.

You send a question and an effort setting
The model reasons with fewer tokens
It writes the final answer
You wait less than with the base model

Good to know: The numbers are the authors' own averages, so your own prompts may differ.

Title: Reasoning saved against the base at xhigh effort
Thinklite-27B: 47.0%
ThinkingCap-Qwen3.8-27B: 45.8%
Swift-1.5-Qwen3.8-27b: 35.6%

• Accuracy: at the low setting it keeps 97.5% of the base accuracy, which is 2.2 points lower
• GGUF: the separate Q4_K_M file cuts reasoning by 48% against an equally quantized base model, while holding 99.3% of that model accuracy
• Other versions: 4 quantized and 0 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the GPU
• Data it reaches: the prompts and images you give it

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 568
• Likes: 51
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 4
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, thinklite, reasoning, efficient-reasoning, qwen3.8, conversational

Links

• Model: https://huggingface.co/Sparset/Thinklite-27B
• Files: https://huggingface.co/Sparset/Thinklite-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Sparset/Thinklite-27B

Then follow the usage in its model card. Nothing in this kit runs by itself.