x feed intelligencenotebook saved October 7

ai · saved October 7

Unsloth trains Qwen3.5 0.8B to 74.3 percent accuracy

@UnslothAI3 sources2 min readlow confidenceoriginal post

1 supported1 contested3 unchecked

Unsloth AI reports raising Qwen3.5 0.8B decision accuracy from 20.7 to 74.3 percent on 4GB VRAM using its open-source repo.

Unsloth AI announced local training of a Decision model named Jev, claiming Qwen3.5 0.8B aggregate accuracy rose from 20.7% to 74.3% across 3 decision benchmarks on 4GB VRAM. The post also names Qwen3.8 and Gemma 4 as convertible LLMs. The linked evidence confirms Unsloth Desktop exists for Mac, Windows, and Linux and that the GitHub repo supports Qwen3.8 and Gemma 4, but none of the three sources contain the benchmark numbers or the Jev model reference.

What it asserts

  1. Qwen3.5 0.8B aggregate accuracy increased from 20.7% to 74.3% across 3 decision benchmarks
  2. Training ran on 4GB VRAM
  3. Any LLM such as Qwen3.8 or Gemma 4 can be turned into a decision model via the Unsloth repo
  4. The repo is open-source
  5. A Decision model named Jev can be trained locally

Evidence

unverifiable
Qwen3.5 0.8B aggregate accuracy increased from 20.7% to 74.3% across 3 decision benchmarks

None of the three sources mention Qwen3.5, the 20.7% or 74.3% figures, or any decision benchmarks. The numbers appear only in the post itself.

Unsloth - Run and Train Models Locally New ✨Introducing Unsloth Desktop
source 1
unverifiable
Training ran on 4GB VRAM

The sources do not reference a 4GB VRAM requirement. Source 1 cites an NVIDIA B200 for video generation, not a 4GB training setup.

On an NVIDIA B200, MiniMax-H3 produced a 960×544, 124-frame video in 13 seconds at 8 steps
source 1
partly supported
Any LLM such as Qwen3.8 or Gemma 4 can be turned into a decision model via the Unsloth repo

The GitHub repo description lists Qwen3.8 and Gemma 4 as supported models, confirming they are supported by Unsloth, but the sources do not describe a decision-model conversion process.

Supports GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, FLUX and more
source 2
supported
The repo is open-source

Source 1 explicitly labels the desktop app open-source and free, and source 2 is a public GitHub repository with 11,519 commits.

Open-source. Free. 100% Local.
source 1 source 2
unverifiable
A Decision model named Jev can be trained locally

No source mentions a model called Jev. The name appears only in the post.

Download for Mac Apple Silicon and Intel .dmg Download for Windows Windows 10 and later .exe
source 1

Why it matters

If the 74.3% figure holds, a 0.8B parameter model on 4GB VRAM could run decision logic on edge hardware without cloud inference costs. Until the benchmark set and evaluation protocol are published, teams evaluating Unsloth for agent or trading pipelines should treat the accuracy claim as marketing.

Keep

  • Unsloth Desktop ships for Mac, Windows, and Linux and supports Qwen3.8, Gemma 4, DeepSeek-V4, MiniMax-H3, and FLUX
  • The GitHub repo has 11,519 commits and is publicly accessible
  • The 20.7-to-74.3% accuracy jump and the Jev model name are not corroborated by any provided source
  • The post links no article or paper; the only linked URLs are the Unsloth homepage, GitHub, and Hugging Face org page

Next

  • Check the Unsloth GitHub repo for any benchmark scripts or eval configs that reproduce the 74.3% figure
  • Review the Hugging Face org page for a Qwen3.5 0.8B model card with training details

Still open

  • Which 3 decision benchmarks were used and what is the evaluation protocol?
  • What is the Jev model and is it published anywhere outside the post?

Sources

  1. unsloth.aiunsloth.ai
  2. github.comgithub.com
  3. huggingface.cohuggingface.co
How this was researched
Unsloth AI Qwen3.5 0.8B decision model training
Unsloth AI Jev decision model benchmarks
unslothqwen3.5local-trainingdecision-modelopen-source

All briefs