Qwen3.8-27B — KERN
KERN uses per-tensor quantization to offer four size and quality tiers. eq denotes an approximate tier: individual tensors use different quantization formats. All files include Q3_K MTP weights.
| Quant | GB | Mean KLD ↓ | 28Q (% of BF16) |
|---|---|---|---|
| KERN-eqQ4 | 14.730 | 0.030233 | 90.48% |
| KERN-eqQ3_M | 13.044 | 0.046237 | 85.71% |
| KERN-eqQ3_S | 11.591 | 0.080428 | 71.43% |
| KERN-eqIQ2_M | 9.580 | 0.164745 | 80.95% |
Scores are relative to BF16’s 28Q result (BF16 = 100%), using the lower primary result where multiple runs exist. The repeat run is excluded. This is a screening metric, not a general capability rating.
KERN-eqQ3_S offers a sub-0.1 mean KLD option at 11.59 GB. KERN-eqIQ2_M is the smallest option at 9.58 GB, with measured mean KLD 0.164745.
Stock comparisons
| Control | GB | Mean KLD ↓ | 28Q (% of BF16) |
|---|---|---|---|
| BF16 | 54.658 | 0 (reference) | 100.00% |
| Stock Q4_K_M | 16.730 | 0.016437 | 85.71% |
| Stock Q3_K_S | 12.257 | 0.088238 | 80.95% |
| Stock IQ2_M | 10.187 | ≈0.1551 | 85.71% |
Usage
Use a llama.cpp build with Qwen MTP support:
./llama-server -m Qwen3.8-27B-KERN-eqQ3_S-MTP.gguf -ngl 99 -c 32768 -fa on -ctk q8_0 -ctv q8_0 --jinja --spec-type draft-mtp --spec-draft-n-max 3
File size excludes KV cache and runtime overhead.
Verify downloads with sha256sum -c SHA256SUMS. File identities are in release-manifest.json, with tensor maps in recipes.