~/LLM/user-benchmarks-inkling-small-276b-12b-against-qwen3-6-27b-on-complex

User Benchmarks Inkling-Small-276B-12B Against Qwen3.6-27B on Complex Coding Task

A Reddit user compared the reasoning and coding performance of Thinking Machines' Inkling-Small-276B-12B model against Qwen3.6-27B using a complex prompt. The test showed Qwen3.6-27B executing structured planning and self-correction within 38 seconds, while Inkling-Small (quantized to UD-Q2_K_XL) thought for 6 minutes and produced less structured code. This comparison highlights the practical trade-offs between different local LLM architectures, showing how quantization levels and reasoning "effort" settings impact real-world coding performance and generation times. It also demonstrates the growing capability of smaller models like Qwen3.6-27B to perform complex, multi-step software engineering tasks efficiently. Inkling-Small was run using Unsloth's UD-Q2_K_XL 2-bit quantization on a DGX Spark GB10 system, where it spent 6 minutes in "max" reasoning effort before outputting code. In contrast, Qwen3.6-27B completed the task much faster, performing multiple self-review cycles, syntax checks, and feature verifications before finalizing its output.

## BACKGROUND

Inkling-Small is a multimodal open-weights model developed by Thinking Machines Labs, featuring 276 billion total parameters but only 12 billion active parameters. Quantization methods like UD-Q2_K_XL compress large models to 2-bit precision to fit on consumer or smaller enterprise hardware, though this can sometimes degrade reasoning quality compared to full-precision or larger active-parameter models.

## REFERENCES

## KEYWORDS

#LLM#Benchmarking#LocalLLaMA#AI Reasoning

$ subscribe --daily

User Benchmarks Inkling-Small-276B-12B Against Qwen3.6-27B on Complex Coding Task | Daily News