---
title: ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw
canonical_url: "https://www.modelscope.ai/models/WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw"
md_url: "https://www.modelscope.ai/models/WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw.md"
repository: WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw
last_updated: 2026-10-12
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model:
  - bottlecapai/ThinkingCap-Qwen3.8-27B
base_model_relation: quantized
parameters: 8.2B
tensor_type:
  - F16
  - I16
  - I32
  - BF16
library_name:
  - safetensors
language:
  - en
downloads: 4
stars: 0
tags:
  - thinkingcap
  - qwen3.8
  - qwen3_5
  - exl3
  - quantized
  - kearuga
  - tensorfold
  - dflash2
  - dgx-spark
  - gb10
  - sm121
---

# ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw

> ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw - An open-source model by WhiteMage on ModelScope. Exact Hugging Face mirror of 0xWhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw. PolyForm Small Business 1.0.0 plus BottleCap personal-use permission; see…

WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw is a 8.2B-parameter text-generation model on ModelScope. licensed under other. derived from bottlecapai/ThinkingCap-Qwen3.8-27B.

- **Repository**: WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 8.2B
- **Base model**: bottlecapai/ThinkingCap-Qwen3.8-27B
- **Tags**: thinkingcap, qwen3.8, qwen3_5, exl3, quantized, kearuga, tensorfold, dflash2, dgx-spark, gb10, sm121
- **Downloads**: 4
- **Stars**: 0
- **Last updated**: 2026-10-12

Source: https://www.modelscope.ai/models/WhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw

---

<img src="assets/header.png" alt="Kearuga — ThinkingCap Qwen3.8 27B. Compact weights. Low divergence. Local reasoning." width="100%">

# Kearuga · ThinkingCap Qwen3.8 27B

**ThinkingCap reasoning. Compact EXL3 weights. Faster shared serving on DGX Spark.**

Kearuga is a **16.35 GB EXL3 release** of [BottleCap AI's ThinkingCap-Qwen3.8-27B](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B), built on [Alibaba Cloud / Qwen Team's Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). It is intended for local reasoning assistants, coding workflows and tool-enabled applications where both fidelity and serving throughput matter.

| Weight download | Mean KLD vs ThinkingCap BF16 | Single-request net decode | 16-request aggregate throughput |
|:---:|:---:|:---:|:---:|
| **16.35 GB** | **0.00517** | **60.61 tok/s** | **249.55 tok/s** |

Speed was measured on **one NVIDIA DGX Spark / GB10** with a TensorFold-based CUDA stack and stock DFlash 2. The 16-request result is a **greedy descriptive measurement**, not single-user speed. KLD is a coarsened held-out measurement, not an accuracy percentage.

[Quality](#quality-and-kld) · [Serving speed](#serving-speed) · [Task timing](#reasoning-task-timing) · [Thinking tokens](#thinking-token-efficiency) · [Download and use](#download-and-use) · [Credits](#credits) · [License](#license)

## Why use this release?

- **Keep the reference model close.** The measured release retains **97.79% top-1 token agreement** with ThinkingCap BF16 on the held-out fidelity suite.
- **Serve more work on one Spark.** The measured serving stack delivers about **37% more aggregate throughput at 16 concurrent requests** than the previous stack using the same target and drafter. This is a workload-specific, descriptive result.
- **Use ThinkingCap's efficient reasoning lineage.** BottleCap reports substantially shorter thinking traces than base Qwen. Those upstream results are credited and separated from this release's measurements below.

**Format:** EXL3, nominal 4.00 decoder bpw. **License:** PolyForm Small Business 1.0.0 plus BottleCap's personal-use permission; this is not an Apache-only derivative.

## Quality and KLD

How closely does the served quantization follow its BF16 reference? The internal held-out fidelity test covers **240 prompts and 25,780 token positions**, spanning chat, code, mathematics, instruction following, tools and long text.

| Metric | Kearuga EXL3 | How to read it |
|---|---:|---|
| Mean coarsened KL divergence | **0.00517361** | Lower is closer to the ThinkingCap BF16 reference |
| Pooled per-position p99 KL | **0.06194665** | A tail-divergence measure, not just an average |
| Top-1 token agreement | **97.7863%** | Agreement at teacher-forced test positions; not task accuracy |
| Tool-domain mean KL | **0.00538761** | Fidelity on the suite's tool slice, not a full tool benchmark |
| Maximum per-position KL | **1.88004674** | Some individual positions still differ substantially |

<details>
<summary><strong>What this fidelity test does—and does not—measure</strong></summary>

The reference is **ThinkingCap BF16**, not unmodified Qwen BF16. The test uses frozen teacher-forced continuations, fixed reference top-20 token support and an other-token bucket. These are **coarsened KL** values, not full-vocabulary KL. The suite is internal; its raw prompts and captures are not included in this model repository.

Low KL does not mean lossless BF16 reproduction or identical answers on every prompt. KL values from other model cards are not directly comparable unless the reference, prompts and measurement method match.

</details>

## Serving speed

The comparison below changes the **serving stack**, not the target model or stock drafter. Both arms use this same quantized target. Measurements were recorded on **10–11 October 2026**, on one DGX Spark / GB10 with a 262,144-token context configuration, BF16 target KV and a 16-request serving limit.

<img src="assets/serving-performance.png" alt="Serving-stack comparison on one DGX Spark: single-request net decode 59.43 to 60.61 tok/s; 4-request aggregate 124.76 to 137.59; 8-request aggregate 167.98 to 214.88; 16-request aggregate 181.91 to 249.55. The 8- and 16-request results are greedy descriptive measurements." width="100%">

| Workload | Previous serving stack | Measured Kearuga stack |
|---|---:|---:|
| Single request: net decode | 59.43 tok/s | **60.61 tok/s** |
| 4 concurrent requests: aggregate | 124.76 tok/s | **137.59 tok/s** |
| 8 concurrent requests: aggregate, greedy descriptive | 167.98 tok/s | **214.88 tok/s** |
| 16 concurrent requests: aggregate, greedy descriptive | 181.91 tok/s | **249.55 tok/s** |
| Cold long-prompt median time to first token | 10.535 s | **7.448 s** |

That is approximately **10% more throughput at four requests, 28% at eight, 37% at sixteen**, and **29% lower cold long-prompt TTFT** on the recorded workloads. These gains belong to the complete serving stack—not quantization alone or an isolated patch.

<details>
<summary><strong>Benchmark conditions and metric definitions</strong></summary>

Single-request and four-request results use a 50-prompt battery across five domains, two runs per boot and two boots per stack. Thinking is enabled at `reasoning_effort=xhigh`, with model-default sampling and a 1,024-token output cap. The eight- and sixteen-request cells use a greedy screen; they are descriptive, not powered claims of universal superiority.

Single-request **net decode** excludes time to first token. Concurrent **aggregate throughput** is total completion tokens divided by the cell's wall time; it includes request overhead and is not a per-user decode rate. Output includes reasoning where enabled. Capped requests are speed probes, not proof that a task was completed correctly.

The baseline is the previous TensorFold-based deployment using the same target and stock DFlash 2. This is not an EXL3-versus-BF16, TensorFold-versus-SGLang, or cross-hardware comparison. Performance depends on prompts, output length, sampling, context occupancy and load. The measured runtime is separate from this weights download.

</details>

## Reasoning-task timing

Token rate is only part of the wait. The table below reports **actual end-to-end wall time per benchmark attempt**, including reasoning and answer generation, on the internal Hard-V2-R suite. These are not estimates obtained by dividing a token budget by tok/s.

<img src="assets/task-latency.png" alt="Descriptive mean task-attempt wall times at four concurrent requests: MATH-500 Level-5 sample 46.10 to 40.17 seconds; AIME 2025 301.74 to 291.83 seconds; combined suite 155.66 to 148.02 seconds. Incorrect and capped attempts remain included." width="100%">

| Reasoning workload | Previous mean time | Current mean time | Observed change |
|---|---:|---:|---:|
| MATH-500 Level-5 sample, 40 problems | 46.10 s | **40.17 s** | **12.9% lower** |
| AIME 2025, 30 problems | 301.74 s | **291.83 s** | **3.3% lower** |
| Combined 70-problem suite | 155.66 s | **148.02 s** | **4.9% lower** |

**Quality alongside timing:** three repeats averaged **62/70 correct**, versus **60/70** for the previous stack on the same target. The registered quality verdict was **non-inferior**, not proof of general superiority.

Timing uses four concurrent requests, thinking at `xhigh` and model-default sampling. Each figure is the equal-weight mean of three existing per-run means. **Incorrect and capped attempts remain included**; these are not successful-task-only completion times. The current AIME runs hit the output cap on 5–6 of 30 attempts; the MATH sample had no cap hits. The observed timing changes are descriptive, not a separately qualified latency win or a guarantee for arbitrary tasks.

## Thinking-token efficiency

**The shorter-thinking contribution comes from BottleCap AI's ThinkingCap model.** In its published BF16 evaluation, BottleCap reports **37.2% fewer reasoning tokens on a macro-average basis**, with average accuracy of **85.8% versus base Qwen's 86.6%**.

The following are selected **publisher-reported upstream results**, **not measurements of this Kearuga EXL3 release**:

| Upstream benchmark | Base Qwen accuracy | ThinkingCap accuracy | Mean thinking tokens: base → ThinkingCap | Fewer thinking tokens |
|---|---:|---:|---:|---:|
| GPQA-Diamond | 89.93% | 88.04% | 12,772 → 7,267 | 43.1% |
| AIME 2026 | 98.13% | 94.27% | 15,663 → 10,934 | 30.2% |
| LiveCodeBench v6 | 91.14% | 91.21% | 28,395 → 22,645 | 20.3% |
| AA-LCR | 81.75% | 84.00% | 2,550 → 1,565 | 38.6% |
| IFBench | 79.75% | 79.71% | 7,961 → 4,266 | 46.4% |

Source: [BottleCap AI's model card at the reviewed revision](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B/blob/7891e03dbda2e9b63a5b9610f20e83ea441f8e2c/README.md). Its evaluation uses an H200 and vLLM, with thinking at `xhigh`; the original card provides the complete benchmark set, uncertainty intervals and methodology. The 37.2% macro reduction is an equal-weight mean of per-benchmark reductions, not a ratio of pooled token counts.

We do **not** claim an additional Kearuga-specific thinking-token reduction. On our internal 70-problem suite, the current and previous serving stacks averaged **5,132 and 5,095 reasoning tokens per attempt**, respectively. The upstream efficiency figures must not be multiplied into our serving-speed results to invent an end-to-end speedup.

## Download and use

### Download the weights

Authenticate with an account that has repository access when required, then download with the Hugging Face CLI:

```bash
hf auth login
hf download 0xWhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw \
  --local-dir ./ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw
```

For reproducible deployments, select an immutable revision in **Files and versions** and add `--revision <commit>` to the download. Verify the included [SHA256SUMS](SHA256SUMS) before use. [FILES.json](FILES.json) inventories the release files; [BENCHMARKS.json](BENCHMARKS.json) records the figures and their scope.

### Choose a compatible runtime

This is **EXL3**, not NVFP4, FP8 or GGUF. It needs a runtime that supports both this model family and EXL3. The results above use TensorFold's **Python-based CUDA path** with a separately obtained stock DFlash 2 drafter. The engine and drafter are **not included** in this repository, and the measured engine image is not a published download here.

Do not assume native TensorFold v1.x or the upstream BF16 SGLang/vLLM recipes can load this pack. A fresh runtime installation or alternative backend does not automatically reproduce the recorded performance or quality results.

### Call an already configured server

The example below connects to a compatible, already running local OpenAI-style endpoint; downloading the weights alone does not start a server.

```bash
curl -N http://127.0.0.1:8888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.8 27B Kearuga (ThinkingCap)",
    "messages": [{"role": "user", "content": "Explain Python context managers with a short example."}],
    "stream": true,
    "max_tokens": 8192,
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20,
    "chat_template_kwargs": {"enable_thinking": true, "reasoning_effort": "xhigh"}
  }'
```

Use your server's model name, port and authentication settings. Thinking is returned separately from the final answer where the runtime supports it. `xhigh` is the upstream recommendation; `medium` and `low` trade reasoning effort against latency and must be evaluated on your workload. The output budget includes reasoning, so allow enough tokens to reach a final answer.

### Memory and context

- **16.35 GB is weight storage**, not total runtime memory. KV cache, drafter, buffers and runtime allocations need additional memory. The measured hardware has 128 GB unified memory.
- **4.00bpw is a nominal decoder budget**, not the whole-file average precision or a promise that every tensor is four-bit.
- The serving configuration allows **262,144 context tokens**. This does not establish that sixteen fully populated maximum contexts fit at once.
- A recorded deployment check accepted a **138,019-token prompt**. That verifies admission, not long-context answer or retrieval accuracy. Vision, video and MTP quality are outside this release's text qualification.

## Credits

| Creator / contributor | Contribution |
|---|---|
| **BottleCap AI** | ThinkingCap fine-tune, its token-efficiency work, direct source weights and governing license |
| **Alibaba Cloud / Qwen Team** | Original Qwen model, architecture, tokenizer and upstream materials |
| **Turboderp and ExLlamaV3 contributors** | EXL3 format and conversion tooling |
| **khronnuz** | Community model resources used during preparation and evaluation; not the direct source of the final weights |
| **Ash (`ashhart`) and TensorFold contributors** | Serving engine and existing planning, verification and sampling machinery |
| **BHCC2025 / BHCC Spark Lab** | [PR #260](https://github.com/ashhart/TensorFold/pull/260) kernel contributions used in the measured runtime |
| **Inco AI; z-lab / Z Lab** | DFlash 2 and its stock mirror; original DFlash research and architecture |
| **0xWhiteMage / Kearuga** | This quantized release, evaluation and serving integration |
| **NVIDIA, PyTorch and Triton contributors** | Hardware and software toolchain used for evaluation and serving |

This release is a quantization of ThinkingCap, not a new ThinkingCap fine-tune. Community engine work is credited to its authors, not claimed as Kearuga-authored kernels. See [ATTRIBUTIONS.md](ATTRIBUTIONS.md) for additional attribution and citations. No contributor endorsement is implied.

## License

The weights retain **PolyForm Small Business License 1.0.0** and BottleCap AI's additional permission for **individual personal, non-commercial use**. Commercial use outside the license's Small Business definition requires a separate license from BottleCap AI. Read [LICENSE](LICENSE) and the [full PolyForm terms](LICENSE-PolyForm-Small-Business-1.0.0.txt) for the actual conditions and definition.

Qwen's upstream materials retain their [Apache-2.0 license](LICENSE-Apache-2.0-Qwen.txt) and [NOTICE](NOTICE). This does **not** make the complete ThinkingCap derivative Apache-2.0. External software and drafter licenses remain separate.

Required Notice: Copyright 2026 BottleCap AI (https://bottlecapai.com)

## Related models and citation

- [Original Kearuga target](https://huggingface.co/0xWhiteMage/Qwen3.8-27B-Kearuga) and [Kearuga DFlash 2 drafter](https://huggingface.co/0xWhiteMage/Qwen3.8-27B-Kearuga-DFlash2): separate NVFP4/FP8 artifacts, not the target/drafter pairing benchmarked on this page.
- [ThinkingCap source](https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B) · [Original Qwen model](https://huggingface.co/Qwen/Qwen3.8-27B) · [Release provenance](RELEASE_PROVENANCE.json).
- Maintainer: [The White Mage](https://benjaminang.tv) · [@0xWhiteMage on X](https://x.com/0xWhiteMage) · [Support on Ko-fi](https://ko-fi.com/0xwhitemage).

```bibtex
@misc{kearuga_thinkingcap_exl3_2026,
  title = {Kearuga: ThinkingCap Qwen3.8-27B EXL3 4.00bpw Release},
  author = {{0xWhiteMage}},
  year = {2026},
  url = {https://huggingface.co/0xWhiteMage/ThinkingCap-Qwen3.8-27B-Kearuga-EXL3-4.00bpw}
}
```

Please also credit the original ThinkingCap, Qwen, EXL3 and DFlash creators when discussing their work.