ThinkingCap-Qwen3.6-27B TQ3_4S — GGUF
TQ3_4S quant of bottlecapai/ThinkingCap-Qwen3.6-27B (Apache 2.0), produced via turbo-tan's llama.cpp-tq3 CUDA fork. ThinkingCap is a Qwen3.6-27B fine-tune optimized for reasoning-heavy tasks. Benchmarked on a Blackwell RTX 5060 Ti (16 GB). Mean AEON score: 0.480 (75 cases, 6144 budget).
File
| File | Size | Quant | BPW |
|---|
ThinkingCap-Qwen3.6-27B-TQ3_4S.gguf | 14 GB | TQ3_4S (turbo-tan) | ~3.8 bpw |
NOT a stock llama.cpp quant
TQ3_4S (TurboQuant 3-bit 4-state) is a custom weight format unique to
turbo-tan/llama.cpp-tq3. Stock llama.cpp and nixpkgs
llama-cpp will exit with
unknown quantization at load time. Use the
llama-server/
llama-cli from the tq3 fork.
Scope of these benchmarks — read this first
These numbers are a light baseline, not a thorough TQ3 evaluation. The mesh's bench framework is built for production agent workload regression-detection on the local stack, not for the kind of multi-axis sweep that upstream quant maintainers typically publish. Specifically:
- Harness scope is bounded. The numbers below come from the mesh's Aeon-Bench-Pod (self-reported mode, 75-case
--fast shape, text-only, no agentic harness). That's a regression suite, not a quality benchmark.
- Sample sizes are small. 75 cases on a single GPU, single rep. None are powered for statistical significance.
- No perplexity / wikitext / MMLU / GSM8K. The mesh's stack isn't a quality benchmark — those are upstream's territory.
- Single GPU class (Blackwell 16 GB). All measurements are on an NVIDIA RTX 5060 Ti 16 GB (CUDA 13.2, turbo-tan/llama.cpp-tq3 77dd77473). No RDNA4, no multi-GPU, no Vulkan. Cross-hardware generalization is NOT implied.
- No human eval. "0.480 mean on the AEON suite" is not a quality verdict on this specific quant.
- 30 prose cases unscored — no frontier judge endpoint configured, so the prose category is excluded from the mean.
What this IS good for: a quick signal that the quant (a) loads, (b) runs at sane throughput, (c) doesn't break the mesh's agent tool-calling. What this is NOT good for: claiming "this is the best quant of this model," reproducing academic benchmark results, or substituting for upstream's validation work.
For a rigorous view, see
bottlecapai/ThinkingCap-Qwen3.6-27B (parent model) and
turbo-tan/llama.cpp-tq3 (quantizer). Raw bench reports are attached as
BENCH-*.md files in this repo.
What we measured
AEON v3 — 75 cases (self-reported, 6144 budget)
Benchmarked on Blackwell RTX 5060 Ti 16 GB, --limit 75 --fast --max-tokens 6144.
| Category | Mean | N |
|---|
| Overall | 0.480 | 75 |
| math | 0.400 | 30 |
| instruction | 0.300 | 30 |
| reasoning | 1.000 | 15 |
| Difficulty | Mean | N |
|---|
| easy | 1.000 | 6 |
| medium | 0.778 | 9 |
| hard | 0.733 | 15 |
| expert | 0.429 | 21 |
| frontier | 0.125 | 24 |
ThinkingCap's profile: pure reasoning specialist (1.000 — perfect on reasoning). Weak on instruction (0.300) and math (0.400). Best suited for roles where reasoning depth is the priority.
Comparison vs. sibling quants (75-case @6144, same harness, node-d)
| Model | mean | math | instruction | reasoning |
|---|
| Tess-4-27B-TQ3_4S | 0.560 | 0.500 | 0.333 | 0.933 |
| ThinkingCap-Qwen3.6-27B-TQ3_4S | 0.480 | 0.400 | 0.300 | 1.000 |
| Qwen3.6-27B-TQ3_4S (prod baseline) | 0.520 | 0.500 | 0.333 | 0.933 |
Quick start
1# Build turbo-tan's tq3 fork
2git clone https://github.com/turbo-tan/llama.cpp-tq3
3cd llama.cpp-tq3
4mkdir build && cd build
5cmake .. -DGGML_CUDA=ON -DCUDA_DOCKER_ARCH=sm_120
6make -j$(nproc)
7
8# Serve
9llama-server \
10 -m ThinkingCap-Qwen3.6-27B-TQ3_4S.gguf \
11 --host 0.0.0.0 --port 8081 \
12 -ngl 99 -c 131072 -t 12 \
13 -ctk q4_0 -ctv q4_0 \
14 -np 1 --batch-size 512 --ubatch-size 128 \
15 --jinja --metrics -rea off
Reproduce the quant
1# Requires the tq3 fork and the F16 source GGUF
2llama-quantize --allow-requantize ThinkingCap-Qwen3.6-27B-f16.gguf \
3 ThinkingCap-Qwen3.6-27B-TQ3_4S.gguf TQ3_4S
Files in this repo
| File | Description |
|---|
ThinkingCap-Qwen3.6-27B-TQ3_4S.gguf | The quantized model (LFS-tracked) |
README.md | This model card |
BENCH-aeon-75-case-6144.md | AEON bench results (75 cases, 6144 budget) |
What's NOT in this repo (caveats)
- Stock llama.cpp will not load this file. TQ3_4S is a custom weight format unique to turbo-tan/llama.cpp-tq3.
- No AMD GPU bench. All measurements are RTX 5060 Ti (CUDA). No RDNA4, no ROCm.
- No quality benchmark (perplexity, MMLU, GSM8K). The custom 3-bit quant works on the mesh's regression tests; whether it's "the best TQ3 quant" needs upstream validation.
- No MTP / speculative-decode bench. ThinkingCap was tested without MTP.
- No prose/creativity evaluation. 30 prose cases were unscored (no judge endpoint configured).
Provenance
- Source model: bottlecapai/ThinkingCap-Qwen3.6-27B — Qwen3.6-27B fine-tune
- Source model license: Apache 2.0
- Quantizer: turbo-tan/llama.cpp-tq3 @ 77dd77473
- Quantizer license: MIT
- Build hardware: NVIDIA RTX 5060 Ti 16 GB (Blackwell), CUDA 13.2, NixOS 25.11
- Bench harness: Aeon-Bench-Pod v1 (self-reported mode)
- Original bench report:
raw/benchmarks/2026-07-09-node-d-thinkingcap-tq3/ in the meshina repo
License
- ThinkingCap-Qwen3.6-27B is Apache 2.0 (per its HF model card).
- turbo-tan/llama.cpp-tq3 is MIT.
- The GGUF in this repo is a derivative of the Apache 2.0-licensed parent, produced with the MIT-licensed quantizer. The Apache 2.0 license is preserved.
