UkisAI Squeezes Swift 1.5's 177B Reasoning Model Into 66 GB

UkisAI has released four mixed-precision GGUF quantizations of Swift 1.5 Qwen3.8-Flash-Next GSQ-RCO, a 177-billion-parameter reasoning model derived from Qwen’s mixture-of-experts architecture. The GGUF release targets local inference through runtimes based on llama.cpp.

According to the upstream Swift Flash Next model card, its post-training reduces thinking tokens by 63.4% and produces a 1.8× speedup while keeping accuracy loss below 1% versus the base model at the xhigh reasoning setting. The GGUF evaluation focuses on quantization drift rather than reproducing those task-level results. A mixture-of-experts model routes each token through a subset of expert blocks, reducing compute relative to a dense model with the same total parameter count. Every expert weight still occupies storage, which explains the large files.

Four tiers, two shards apiece

Each tier arrives as two GGUF shards whose combined size appears below. Both shards must remain in the same directory; loading shard one allows llama.cpp to discover shard two automatically. A separate BF16 vision projector handles image-text input and adds to the storage and memory requirements.

Kullback-Leibler divergence, or KLD, measures the difference between the quantized model’s next-token distribution and a BF16 reference. Lower values indicate less distribution drift. Coding accuracy, tool use, multilingual task performance, long-context stability, and vision quality require separate benchmarks.

IQ2_XS leads near 68 GB

At 68.15 GB, IQ2_XS offers the strongest reported size-to-drift result among the smallest practical tiers. Its KLD is 5% to 11% lower than ISTA-DASLab’s GSQ-RCO IQ2_XS on seven of eight reporting sets. The model card reports reductions of 8.9% on mathematics text and 10.5% on Chinese text. Compared with its own pre-refinement Swift quant, it lowers KLD across every reported domain by 8% to 17%.