Qwen3.8 27B Kearuga (ThinkingCap)
A 16.35 GB EXL3 quant of BottleCap's ThinkingCap-Qwen3.8-27B reasoning model, served on a single DGX Spark.
Overview
The third entry in the Kearuga collection: an EXL3, nominal 4.00bpw release of BottleCap AI's ThinkingCap-Qwen3.8-27B, a reasoning model built on Qwen3.8-27B. It is aimed at local reasoning assistants, coding workflows and tool-enabled applications where fidelity and serving throughput both matter.
Where the original Kearuga suite tuned a target-and-drafter pair for SGLang, this release asks a different question: how closely can a compact weight file track a reasoning reference while still serving real concurrency on one Spark?
TL;DR
- 16.35 GB EXL3 weights, nominal 4.00 decoder bpw
- 97.79% top-1 token agreement vs ThinkingCap BF16, 0.00517 mean KLD
- 60.61 tok/s single-request decode, 249.55 tok/s at 16 concurrent
- TensorFold CUDA recipe for DGX Spark, experimental 24 GB GPU path
Fidelity
Measured against the ThinkingCap BF16 reference on a held-out suite of 240 prompts and 25,780 token positions spanning chat, code, mathematics, instruction following, tools and long text: 97.79% top-1 token agreement and a mean coarsened KL divergence of 0.00517, with a pooled per-position p99 of 0.0619. The coarsened measurements are documented on the model card rather than presented as accuracy percentages.
Serving
Benchmarked on one DGX Spark (GB10) with a TensorFold-based CUDA stack, stock DFlash 2 and a 262K context configuration with BF16 target KV: 60.61 tok/s single-request decode and 249.55 tok/s aggregate at 16 concurrent requests, roughly 37% more aggregate throughput than the previous stack on the same target and drafter.
The launch recipes, hardware configurations and a documented quick start live in the serving repository.
24 GB GPUs
A 24 GB NVIDIA GPU path exists for Ampere, Ada and Blackwell candidates, documented in the repository. It is experimental with a known loading-admission limitation, and RTX 5090 D V2 and 5090 Laptop variants are candidates rather than qualified targets; all published benchmarks remain Spark-only.
Status
Released October 2026 on Hugging Face and ModelScope, under PolyForm Small Business 1.0.0 plus BottleCap's personal-use permission. The serving stack is on GitHub.