fraQtl/Qwen3.8-27B-Hi-Fi-GGUF overview
Qwen3.8 27B Q4 K M GGUF — fraQtl Hi Fi calibration aware quantization Text only GGUF the vision tower is not converted . Iso size Q4 K M class build of Qwen/Qw…
Runs locally from ~15.33 GB disk (16 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| Qwen3.8-27B-fraQtl-HiFi-Q4_K_M.gguf | GGUF | Q4_K_M | 15.33 GB | Download |
Model Details
| Model ID | fraQtl/Qwen3.8-27B-Hi-Fi-GGUF |
|---|---|
| Author | fraQtl |
| Pipeline | text-generation |
| License | apache-2.0 |
| Base model | Qwen/Qwen3.8-27B |
| Last modified | 2026-08-21T12:09:01.000Z |
Model README
---
license: apache-2.0
base_model: Qwen/Qwen3.8-27B
base_model_relation: quantized
library_name: gguf
pipeline_tag: text-generation
tags:
- qwen
- qwen3.8
- gguf
- q4_k_m
- calibration-aware-quantization
- fraqtl
---
Qwen3.8-27B Q4_K_M GGUF — fraQtl Hi-Fi (calibration-aware quantization)
Text-only GGUF (the vision tower is not converted). Iso-size Q4_K_M-class
build of Qwen/Qwen3.8-27B with fraQtl calibration-aware per-tensor
quantization: byte-matched to the leading community Q4_K_M (within 0.001%),
measurably closer to the original model's output distribution on both locked
evaluation slices.
Evaluation — fidelity to the original model
Metric: symmetric top-20 KLD on the support of a Q8_0 teacher **quantized from
the original pinned bf16 checkpoint** (fidelity-to-original — the axis a user
migrating from the original model experiences; teacher-framing caveat applies).
3 independent runs, run-to-run drift ≤1.4e-17 (byte-deterministic rig), one
shared teacher and one pinned eval slice for every arm.
| Arm | bytes | KLD code/math ↓ | KLD general ↓ |
|---|---|---|---|
| fraQtl Hi-Fi (this repo) | 16,464,274,816 | 0.12297 (−45.3%) | 0.23120 (−18.0%) |
| unsloth UD-Q4_K_M (canonical) | 16,464,440,224 | 0.22493 | 0.28209 |
| ggml-org Q4_K_M (second) | 18,973,870,432 | 0.19074 | 0.24274 |
Margins are vs the canonical arm. Sizes are not iso across all three arms:
this build is byte-matched to the unsloth canonical by construction
(Δ165,408 B, 0.001%); the ggml-org file is a genuinely different, larger
build (18.97 GB, no NextN block — disclosed). Read every row as
fidelity-at-stated-size.
Disclosed losses: Wikitext-2 perplexity is measurably worse for this build
(+0.7% vs canonical) and general-slice top-1 token agreement is marginally
worse (−0.2pp). Both are in the build receipt.
Task benchmarks — ties, stated as ties
GSM8K (n=200) and MATH-500 (n=500), 0-shot instructed-CoT, identical harness
for all three arms:
| Arm | GSM8K | MATH-500 |
|---|---|---|
| fraQtl Hi-Fi | 95.0% (190/200) | 87.6% (438/500) |
| unsloth canonical | 94.0% (188/200) | 88.2% (441/500) |
| ggml-org second | 94.0% (188/200) | 87.4% (437/500) |
**Every pairwise difference is inside the 95% confidence interval (GSM8K
±6.9pp at n=200; MATH-500 ±4.4pp at n=500) — these are statistical ties in
every direction, not wins for any arm.** A per-item paired audit of GSM8K
sharpens this: ours vs canonical differ on only 4/200 items (3–1 split,
exact two-sided sign test p = 0.625 — indistinguishable from a coin flip).
The KLD margins above are the only metrics on this card claimed as wins.
Long-context retrieval (needle-in-a-haystack)
Grid: contexts {8K, 32K} × depths × 3 passkeys, plus 131K cells × 2 passkeys —
63 cells per arm per run, multi-run. **All three arms are perfect on every
cell of every run (270 individual passkey checks total)** — reported as
parity, not a differentiator.
Provenance
| Field | Value |
|---|---|
| Base model | Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 (Apache-2.0) |
| This artifact | Qwen3.8-27B-fraQtl-HiFi-Q4_K_M.gguf, 16,464,274,816 B, sha256 43bd8f235bd7843b6bb06d632235b3e2430c23ad3f54b2fb1e56f35e74eebd8d, true bpw 4.8210, 866 tensors (matches the canonical's tensor set, incl. NextN block) |
| Toolchain | llama.cpp @ 4df29be4f4c3673f428170fda944a5b19f743bb8 — single commit for conversion, quantization, and every eval, all arms. Conversion-correctness gate (bf16-GGUF vs HF reference, hybrid GDN arch): top-1 0.983, KLD 2.4e-4 — PASS |
| Calibration | A1-selector winner mix_cm70_gen30, llama-imatrix at ctx 8192; imatrix.dat (sha256 c36c3d06…3dbd0d) included in this repo |
| Teacher | one Q8_0-from-our-bf16 instance, used identically for all three arms |
| Baselines pinned | unsloth Qwen3.8-27B-GGUF @ 990216cf3125 (322e194f…3482); ggml-org Qwen3.8-27B-GGUF @ 0669b98607d4 (31629f53…3d34) |
| Eval runtime | llama-cpp-python 0.3.35, A100-80GB, seed 0 |
Fine print
- One MATH-500 cell (ours) was accidentally run twice due to a preemption
that looked like a wedge: an isolated re-fire and the original delayed run
both completed independently and converged on the **exact same score
(438/500)** — an incidental determinism receipt. Full detail in the
taskbench receipt.
- 262K is this model's native context; the numbers on this card are a
quality-tier GGUF claim, not a KV/serving claim.
- Receipts for every number are in
receipts/(JSON: pins,
conversion gate, calibration selection, build + 3-run KLD, NIAH grid,
task benches).
Credit
(d4nnyy/AtomicChat): prior day-0 stored-logits KLD across 36 quants
(4096-ctx, no retrieval gate) — credited as prior art; we do not claim
first KLD receipts on this model. Our scope differs: independent teacher,
32K–131K contexts, needle-gated, 3-run.
- unsloth: the canonical community build
(4.3M downloads/30d), UD-Q4_K_M class.
- ggml-org / llama.cpp: toolchain
and second-arm GGUF.
Files
| File | What |
|---|---|
| Qwen3.8-27B-fraQtl-HiFi-Q4_K_M.gguf | the artifact (16.46 GB) |
| imatrix.dat | calibration importance matrix (reproducibility) |
| receipts/ | receipt JSONs for every number on this card |
The fraQtl ladder
| Tier | Job | Example |
|---|---|---|
| Hi-Fi | maximum fidelity at standard 4-bit size | this repo |
| Hi-Fi Phone | fits phone/laptop RAM, fidelity-first | E2B Phone, 2.86 GB |
| SmartEdge | smallest footprint that survives | E2B IQ3XXS, 2.45 GB |
Same discipline at every tier: pinned provenance, measured numbers, losses
disclosed.
More from fraQtl
The serving lane — KV-cache compression sidecars for vLLM — holds
nine concurrent ≈128K-context users on a single A100 (134.1 tok/s
aggregate, 9/9 per-user retrieval checks, receipt 2026-08-14):
fraQtl/qwen3-4b-instruct-2507-kv-sidecars.
Org page: huggingface.co/fraQtl.
Run fraQtl/Qwen3.8-27B-Hi-Fi-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models