Sabi-Noema: A capability and efficiency focused fine-tune of Qwen 3.8 27B

Sabi-Noema and Sabi-Noema-MLA are two post-trained versions of Qwen 3.8 27B designed to push the capability frontier for models under 30B parameters. Trained exclusively on open-source environments, these models achieve a 2x reduction in verbosity while improving performance in coding, agentic tasks, and tool calling.

Benchmark Performance

The models demonstrate superior efficiency and accuracy compared to competitors:

• Coding & Agentic Tasks: Sabi-Noema beats GPT 6 Luna-medium on DeepSWE, placing between Luna 6 medium and high. It also outperforms Luna-high on agentic and tool calling benchmarks including TB2.1, Tau-3 bench, ClawEval, and BFCL v4.
• Hallucination: The models exhibit a considerably lower hallucination rate than Luna, base Qwen, and other competing models.
• Reasoning Efficiency: Sabi-Noema generates 50% fewer tokens per response on TB2.1 and 25% fewer tokens on DeepSWE compared to the base Qwen model.

MLA Conversion and Memory Efficiency

Sabi-Noema-MLA utilizes a post-hoc MLA conversion of Qwen's GQA attention, requiring only a calibration dataset and no extra training.

• KV Cache Reduction: With BF16 latents, the model requires 3.6x lower bytes per token. Compressing MLA latents to FP8 reduces the total KV cache size on TB2.1 by 8.4x with similar performance.
• Context Length: When using q4_0 quantization on a 24GB GPU (such as 3090/4090), Sabi-Noema-MLA supports a context size of ~534K, whereas base Qwen with FP8 cache is limited to ~150K.