APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

TurinTech's Artemis Shrinks Google's Gemma 4 E2B to 2.05 GB Without Custom Code

TurinTech released Gemma 4 E2B Artemis GGUF : three files, 2.05 to 2.31 GB, down from 9.3 GB. The default Q3_K_L file matches the original model's next token 95.8% of the time. Other public quants of the same checkpoint below 2.6 GB lose 23 to 34 task points.

TurinTech's Artemis Shrinks Google's Gemma 4 E2B to 2.05 GB Without Custom Code
Source
AlphaSignal
Published
Author
AlphaSignal
Read
2 min read

Takeaways

  • TurinTech released Gemma 4 E2B Artemis GGUF : three files, 2.05 to 2.31 GB, down from 9.3 GB.
  • The default Q3_K_L file matches the original model's next token 95.8% of the time.
  • Other public quants of the same checkpoint below 2.6 GB lose 23 to 34 task points.
  • The trick: keep Google's QAT Q4_0 weights and compress only the per-layer embedding tables harder.
  • Free under Apache-2.0, text only, runs on stock llama.cpp, Ollama and LM Studio.
  • Weak spots: longer outputs drift, small math losses, no memory measurements yet; Unsloth's 2.62 GB file is more faithful.

TurinTech, the company behind the Artemis code optimisation platform, has released three compressed builds of Google’s smallest Gemma 4 chat model in its Artemis GGUF repository . The 2.05 GB to 2.31 GB files are 75% to 78% smaller than the 9.3 GB reference checkpoint.

All three builds use the standard GGUF format and run with stock llama.cpp , without forks or custom kernels. TurinTech’s benchmarks show them retaining more of the original model’s behaviour than the other public quantizations it tested at similar sizes.

The public baseline drops sharply below 2.6 GB

At 9.3 GB, the reference checkpoint exceeds the physical memory of an 8 GB phone before accounting for the operating system, application or inference overhead. Quantization reduces that footprint by storing model weights with fewer bits, although aggressive compression can change the model’s token choices and reduce task accuracy.

TurinTech benchmarked each file against the reference checkpoint using the same llama.cpp settings on an AMD Ryzen 9 9950X CPU. Unsloth’s 2.62 GB UD-Q4_K_XL remained the closest public build. Other tested builds below that size chose a different next token 45% to 55% of the time and lost 23 to 34 percentage points across the sampled tasks.

The reported next-token metric checks whether a build selects the same highest-probability token as the reference. The 32-token metric compares their generated prefixes, while task accuracy measures outcomes across a 310-item evaluation set.

BuildSizeNext-token agreement32-token match rateTask accuracy change
Unsloth UD-Q4_K_XL2.62 GB98.3%89.1%−0.3 points
Artemis Q3_K_L, default2.31 GB95.8%74.7%−1.6 points
Artemis Q3_K_M2.20 GB93.1%69.0%+0.3 points
Artemis IQ3_XXS2.05 GB91.6%67.1%0.0 points

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01huggingface.co
  2. 02turintech.ai
  3. 03huggingface.co
  4. 04huggingface.co
  5. 05github.com