Xananthium/Calmer-Qwen3.8-27B-4bit

Calmer Qwen 3.8 27B · 4-bit

An experimental fixed-weight edit by Xananthium, studied at medium reasoning and temperature 1.3. The name “Calmer” refers to observed text behavior in specific tests. The study measures completion, repetition, language patterns and correctness separately.

Research and evidence: Qwen3.8-Weight-Space-Programming.

What is in this release

File or directory	Purpose	Validation
Calmer-Qwen3.8-27B-Q4_K_P-F32Edits.gguf	Complete merged model: the original mixed 4-bit-family GGUF with three edited down-projection tensors stored as F32	All other tensor payloads byte-identical; CPU algebra checks and native CPU load/decode smoke passed. Its HLE/GSM performance has not been benchmarked.
calmer_delta.npz	Exact fixed rank-one edit used in the measured serving experiment	Byte-identical to the selected experimental delta
runtime/	Portable source package for the packed GGUF plus fixed FP32 correction representation	CPU arithmetic, tensor-parallel binding, relocation and configuration checks passed. CUDA sources must be built; a rebuilt runtime needs fresh numerical validation.
hf-config/	Pinned configuration, tokenizer, processor and chat template	Copied from the experimental source profile, with file hashes
mmproj-Qwen3.8-27B-BF16.gguf	Unchanged optional vision projector	Separate from the text-model edit
merged_export_manifest.json, release_manifest.json	Identities, tensor changes, precision and provenance	Includes the complete merged tensor inventory

Representation matters. The reported experimental results below were obtained with the original packed GGUF matmul plus a fixed FP32 sidecar correction, followed by the recorded BF16 cast/subtraction order. The merged file computes the same intended weight update in materialized F32 matrices, but a GGUF engine can use different kernels and rounding. Do not attach the historical scores to the merged file as if it had undergone the same evaluation.

The source is a mixed-precision Q4_K_P-family model, not uniformly four-bit weights. Its edited MLP down-projections were Q5_K. In the merged download, only those three matrices are promoted to F32 to avoid an additional requantization of the edit. The filename identifies this explicitly. The merged file is 18,809,113,472 bytes.

Edit definition

All layer indices are zero-based. Every original block executes once; no repeated blocks are enabled.

W_new = W - alpha * outer(u, coeff)
global alpha = 1.5
effective layer strengths: 49 = 0.25, 50 = 1.5, 51 = 0.25
selected input columns per layer = 64

The saved coefficient arrays already encode the relative flank scaling. The update is a fixed linear map; the correction applied to an activation varies with that activation. It does not inspect the prompt or select answers from a lookup table. The direction and sparse support were estimated from constructed productive/repetitive GSM continuations and decoded weights. No gradient optimizer was used, but the intervention is data-fitted.

Observed local results for the tested sidecar representation

These are small local panels, not full benchmark scores. The original and edited conditions used identical frozen requests within each comparison, with medium reasoning, temperature 1.3, top-p 0.95, top-k 20 and repetition penalty 1.

Panel	Original	Selected edit
HLE: fixed 32-question text subset, strict correct	3/32	4/32
GSM8K: 24 previously exposed retention cases	24/24	23/24
Authored state transfer: agent-reviewed intended correct	4/4	3/4
Constructed repetitive continuations: completed correct	1/2	2/2
Generated tokens across those two continuations	4,906	1,228
Four prescribed next-action tool corrections	4/4	4/4

The HLE change comprises three gains and two losses. One gained correct choice has an arithmetic defect in its explanation. None of ten prespecified completed baseline misses was repaired. The selected edit uses 12.8% more total tokens on the 32-question HLE panel, despite much shorter constructed repetition continuations. It also introduces a real GSM error and a real state-tracking error. Broader improvement remains unproven.

Four development HLE questions were also tested at two random seeds, plus a separate planar-graph question: 37 distinct HLE questions in the actual-serving comparison, not the whole HLE exam. Both game/tree answer choices are strictly correct at seed 20261107 for this selected profile, but the game existence proof is invalid and the tree has a valid count certificate followed by false evolutionary-history prose. At seed 20261105, the game proof is valid and the tree remains unfinished. Correct choices and valid explanations are reported separately.

The writeup includes all five actual-serving profiles, fourteen matched laboratory conditions, 503 saved text generations, exact thinking/final token accounting, normalized lexical markers, character degeneration, proof reviews, unsuccessful variants and adaptive-selection limitations. The constructed repetition examples were reused during discovery; they are not independent evidence of real-world agent-loop repair. No generated tool calls were executed.

Download and use

hf download Xananthium/Calmer-Qwen3.8-27B-4bit --local-dir Calmer-Qwen3.8-27B-4bit

Merged GGUF

Use a GGUF engine supporting the qwen35 architecture and mixed tensor types. A CPU load/decode check with llama.cpp commit b21e4de74567f5eef213765c9476a843c2e43f0d generated the correct answer to the supplied simple arithmetic smoke prompt. This validates that the artifact loads and generates; it is not a reasoning benchmark or sidecar-equivalence test.

From the downloaded directory, the tested smoke command is:

llama-completion --model Calmer-Qwen3.8-27B-Q4_K_P-F32Edits.gguf \
  --n-gpu-layers 0 --threads 8 --threads-batch 8 --ctx-size 2048 \
  --n-predict 128 --temp 1.3 --top-p 0.95 --top-k 20 --min-p 0 \
  --repeat-penalty 1 --seed 20261105 --no-conversation \
  --no-display-prompt --no-warmup --file native_smoke_prompt.txt

Use the supplied chat template for normal conversational prompting. A fully quantized re-export is another numeric artifact and requires its own evaluation.

Original packed weights with the tested fixed edit

Obtain the exact upstream base file:

hf download HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF \
  Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --revision 993a5971fda8f30dd1b7eb2654792ba4415c7460 \
  --local-dir upstream-base

The historical setup also requires the pinned separate DFlash draft:

hf download syvai/Qwen3.8-27B-DFlash2-W4A16 \
  --revision 4d30ec736ffc6b8688dc2ae2b502d9b48bdec279 \
  --local-dir dflash-draft

Follow runtime/README.txt to provision and build the pinned custom vLLM/GGUF stack. The portable package includes the required source overrides and CUDA source closure, with licensing and build constraints. The tested binary is not redistributed; rebuilding it is not a claim of token-identical reproduction. From the downloaded model directory, with upstream-base/ and dflash-draft/ downloaded inside it, verify the runtime using:

python -B runtime/cpu_checks.py
python -B runtime/launch_exact.py \
  --model-path upstream-base/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf \
  --delta-path calmer_delta.npz --hf-config hf-config \
  --mmproj-path mmproj-Qwen3.8-27B-BF16.gguf \
  --draft-dir dflash-draft --state-dir "$HOME/.cache/calmer-qwen38" \
  --check-imports

This preserves the historical TP2, TQ4 KV and DFlash7 configuration. Remove --check-imports to serve after provisioning the matching environment and building the extension. Check both workers' startup audits and validate the rebuilt GPU runtime before comparing scores. Local state and compile caches must remain outside this distributable model directory.

Do not apply the sidecar again to the merged GGUF. That would apply the edit twice. The exact-runtime launcher verifies the upstream base checksum.

Identity and attribution

• Upstream derivative: HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF, revision 993a5971fda8f30dd1b7eb2654792ba4415c7460.
• Base GGUF SHA256: ba36dc3c2b2ff5e0aa5d71092a8894546996a6a119ae391803dda07cdc08516d.
• Exact selected delta SHA256: ebb80ab4f906b7c7fbc61e95dd285b1e27b82b35e3e68982e575ccbfe80f8b51.
• Merged GGUF SHA256: 9875d0d1f9ad37599ecd5b9b9aa5e32b3a585aeff83af64df52a00aceea6cab8.
• The public name follows the upstream Qwen3.8 label; the GGUF architecture is qwen35, and the loader uses Qwen3_5 metadata. These identifiers alone do not establish an upstream training lineage.
• Credit to Qwen and HauhauCS for the source model. This is an experimental derivative by Xananthium, not an official Qwen release. See LICENSE, NOTICE.txt, and the runtime's third-party notices.

The model card, tooling and scientific reviews were prepared with AI assistance. Full benchmark questions and HLE-containing raw traces are excluded from public research exports in accordance with the dataset card's redistribution request; the research repository provides IDs, scores, hashes and authorized reconstruction instructions.