Qwen3.6 27B SuperDense
> "Local intelligence... to the max."
Model source
- Kind
- Text model
- File size
- 11.7 GB
- Base model
- Qwen/Qwen3.6-27B
- Version
- vRevision 23b2394
- Updated
- Oct 2, 2026
- Publisher
- @macwhisperer
- Downloads
- 125
- Parameters
- 27B
- Source
- Hugging Face
Source description
π Qwen3.6-27B-Dense-Imatrix-IQ3_M.gguf (2026 Edition)
"Local intelligence... to the max."
This is a custom-quantized version of Qwen3.6-27B, specifically optimized to obtain the highest possible local byte-intelligence ratio with 24GB+ RAM consumer laptops or computers.
π§ Why this model is different
Unlike a standard quant, this model was processed using a custom Importance Matrix (imatrix). The training data for the imatrix was hand-curated to preserve:
- Incredible reasoning: Inclusion of custom coding examples built with frontier models provides high retention of very specific and sharp architectural reasoning skills
- Logical Flow: Inclusion of
llama.cppsource code, logic puzzles, and historical writing in the imatrix training to ensure the model stays coherent at low bitrates. - High Speed: Built using llama.cpp specifically for local-first AI and edge computing setups like apple silicon with minimum 24GB RAM
π Quantization Details
- Base Model: Qwen3.6-27B
- Quantization: IQ3_M
- Format: GGUF
- Size: ~12.58 GB
- Context Length: 262144 tokens
βοΈ Recommended Inference Settings
Optimize for balance between creativity and coherence:
--repeat-penalty: 1.1 β 1.4 (Sweet spot! Pushes away from familiar loops. >1.5 causes "robot-speak".)--repeat-last-n: 128 β 256 (Larger window ensures the model doesn't forget recent repetitions.)--temperature: 0.7 β 0.8 (Prevents over-committing to safe/repetitive tokens.)--top-p: 0.90 (Trims low-probability hallucinations without killing creativity.)--min-p: 0.05 β 0.1 (Optional: Prunes very low-probability tokens if your backend supports it.)
π Perplexity Benchmarks
The following results were generated using llama-perplexity on the wikitext-2-raw/wiki.test.raw dataset.
| Model | Precision | Perplexity (PPL) | Ξ PPL |
|---|---|---|---|
| Qwen3.6-27B- (no-imatrix) | IQ3_M | 7.4952 | - |
| Qwen3.6-27B- (Imatrix) | IQ3_M | 7.1485 | -0.3467 |
βοΈ Evaluation Verdict
Refined Accuracy: A PPL reduction of -0.3467 indicates that the I-Matrix successfully "pinned" the critical weights for the Gated DeltaNet layers, further smoothing out the IQ3_M experience.
π Hardware Performance (Apple M2)
coming soon
π Links
Check out my other models!
24GB+ (RAM)
16GB+ (RAM)
8GB+ (RAM)
4GB+ (RAM)
All make excellent companions to this model!
Sources
1 source
Sources
1 sourceVerified Oct 2
Model artifacts
1 artifact
Model artifacts
1 artifactQwen3.6-27B-Dense-Imatrix-IQ3_M.gguf
gguf Β· 11.7 GB Β· SHA-256 b23c684de6bdβ¦17d3 Β· Hugging Face
DownloadSource excerpts
3 excerpts
Source excerpts
3 excerpts--- license: apache-2.0 base_model: - Qwen/Qwen3.6-27B tags: - GGUF - Imatrix - Qwen - Dense - Qwen3.6 - LLM - Text --- # π Qwen3.6-27B-Dense-Imatrix-IQ3_M.gguf (2026 Edition) > "Local intelligence... to the max." This is a custom-quantized version of **Qwen3.6-27B**, specifically optimized to obtain the highest possible local byte-intelligence ratio with 24GB+ RAM consumer laptops or computers. ## π§ Why this model is different Unlike a standard quant, this model was processed using a **custom Importance Matrix (imatrix)**. The training data for the imatrix was hand-curated to preserve: * **Incredible reasoning:** Inclusion of custom coding examples built with frontier models provides high retention of very specific and sharp architectural reasoning skills * **Logical Flow:** Inclusion of `llama.cpp` source code, logic puzzles, and historical writing in the imatrix training to ensure the model stays coherent at low bitrates. * **High Speed:** Built using llama.cpp specifically for local-first AI and edge computing setups like apple silicon with minimum 24GB RAM ## π Quantization Details - **Base Model:** Qwen3.6-27B - **Quantization:** IQ3_M - **Format:** GGUF - **Size:** ~12.58 GB - **Context Length:** 262144 tokens ### βοΈ Recommended Inference Settings Optimize for balance between creativity and coherence: * **`--repeat-penalty`: 1.1 β 1.4** *(Sweet spot! Pushes away from familiar loops. >1.5 causes "robot-speak".)* * **`--repeat-last-n`: 128 β 256** *(Larger window ensures the model doesn't forget recent repetitions.)* * **`--temperature`: 0.7 β 0.8** *(Prevents over-committing to safe/repetitive tokens.)* * **`--top-p`: 0.90** *(Trims low-probability hallucinations without killing creativity.)* * **`--min-p`: 0.05 β 0.1** *(Optional: Prunes very low-probability tokens if your backend supports it.)* ### π Perplexity Benchmarks The following results were generated using `llama-perplexity` on the `wikitext-2-raw/wiki.test.raw` dataset. | Model | Precision | Perplexity (PPL) | Ξ PPL | | :--- | :--- | :--- | :--- | | Qwen3.6-27B- (no-imatrix) | IQ3_M | **7.4952** | - | | **Qwen3.6-27B- (Imatrix)** | **IQ3_M** | **7.1485** | **-0.3467** | ### βοΈ Evaluation Verdict Refined Accuracy: A PPL reduction of -0.3467 indicates that the I-Matrix successfully "pinned" the critical weights for the Gated DeltaNet layers, further smoothing out the IQ3_M ex...
Source context: 125 downloads Β· 4 likes Β· Repo macwhisperer/Qwen3.6-27B-SuperDense
Source context: 490 downloads Β· 3 likes Β· Repo macwhisperer/Qwen3.6-27B-SuperDense
