Xananthium/Calmer-V2-Qwen3.8-27B-4bit

Calmer V2 · Qwen3.8 27B · 4-bit

A small, fixed early-layer edit on top of Calmer V1. Selected by its owner for daily use after a promising efficiency pilot.

Research and contribution repository · V2 report and thinking examples · Calmer V1

V2 adds a learned sparse rank-one correction at zero-based layer 18, while preserving Calmer V1's 49/50/51 corrections byte for byte. It uses medium reasoning and temperature 1.3. In the exploratory five-question panel, all five final answers stayed correct while generated tokens fell 22.45%. One additional topology answer used 58.19% fewer tokens, and a deliberately primed arithmetic recheck fell from 15 repetitions to one.

The owner selected the pilot for production and reported a favorable experience after using it. That is qualitative feedback, separate from the recorded development measurements.

What to download

Artifact	Purpose	Validation scope
Calmer-V2-Qwen3.8-27B-Q4_K_P-F32Edits.gguf	Complete merged model; mixed-Q4 base with four edited matrices stored as F32	Tensor-preservation, algebra and native CPU load/decode checks passed. No HLE/GSM evaluation of this representation
calmer_delta.npz + runtime/	Exact correction weights and portable source for the packed-base/FP32-correction approach	This is the representation used by the measured and live pilot; a rebuilt custom runtime requires its own numerical validation
mmproj-Qwen3.8-27B-BF16.gguf	Unchanged upstream vision projector	Included with its pinned identity; V2 experiments were text-only
hf-config/	Pinned tokenizer, configuration and chat template	Same assets as the serving pilot
release/, SHA256SUMS	Export checks, smoke evidence and artifact identities	Distinguishes merged export from sidecar results

The original packed base is downloaded separately for the sidecar runtime. Apply the delta only to that original base, never to the merged V2 GGUF. The merged file already contains the edits.

Observed V2 results

These are previously exposed, adaptively used development questions, not full benchmark scores. Comparison conditions, serving-history differences and all failed variants are documented in the V2 report.

Measurement	Calmer V1 reference	V2 pilot
Fast panel: two difficult questions + three GSM8K checks	5/5 final answers; 23,831 tokens	5/5 final answers; 18,481 tokens
Game explanation	Valid general proof; 12,743 tokens	Valid general proof; 8,177 tokens
Tree explanation	Correct choice, flawed event history	Correct choice, additional event-history errors
Topology development question	Correct answer, sound core; 25,467 tokens	Correct answer, sound core; 10,648 tokens
Primed cookie continuation	Correct; 862 tokens; 15 exact rechecks	Correct; 554 tokens; one exact recheck
Six-case content review	6/6	6/6; strict parser grade 5/6
Hat puzzle	Wrong completed answer	Unfinished at 32,768 tokens
Wider 24-case GSM8K check in this follow-up	24/24	Not run for V2

The topology answers retain mathematical display/sign errors despite a sound core argument. Shorter thinking did not consistently mean cleaner reasoning or faster completion of the entire concurrent batch. The towel continuation still repeated six times.

A GPU driver fault prevented the planned warm-start and second-seed confirmation and V2's wider GSM check. The machine subsequently recovered, and the exact sidecar pilot passed production startup, both-worker patch audits, and public API smoke checks. Those operational checks do not replace the missing benchmark confirmation.

The follow-up explored 14 new weight variants, saved 248 response records and observed 933,411 generated tokens, including an explicitly interrupted failure. These counts include repeated questions. No general HLE or GSM8K improvement is established.

Use the merged GGUF

hf download Xananthium/Calmer-V2-Qwen3.8-27B-4bit \
  Calmer-V2-Qwen3.8-27B-Q4_K_P-F32Edits.gguf \
  --local-dir calmer-v2

Load it in an engine compatible with the source checkpoint's qwen35 GGUF architecture and chat template. Recommended starting sampling settings are temperature 1.3, top-p 0.95, top-k 20, repetition penalty 1, with medium reasoning where the engine supports that template setting. Keep a generous output budget for difficult questions; short smoke-test limits are not the evaluation budgets.

The pinned hf-config/generation_config.json retains upstream defaults for provenance; the V2 launcher overrides them. Other engines should explicitly use the settings in recommended_generation_config.json and medium reasoning.

The published native CPU smoke is a load/decode check, not a claim that every GGUF frontend supports the model or reproduces the custom runtime's behavior.

Reproduce the serving representation

Download the repository and follow runtime/README.txt. It includes pinned original-base and DFlash-draft download commands, CPU preparation, custom CUDA/vLLM requirements, and the launcher.

The measured local stack preserves:

• DFlash with seven speculative tokens and its separate W4A16 draft.
• TurboQuant four-bit target-model KV cache and native 80 GiB RAM KV offloading.
• 95% GPU memory budget, tensor parallelism across two GPUs, BF16 computation.
• 200,000-token context, 16 concurrent sequences, 4,096 batched tokens.
• Compiled execution with decode CUDA graphs and prefix caching.
• Medium reasoning, temperature 1.3; up to five images per request.

These are runtime features, not features baked into the GGUF file. Stock vLLM is not the pinned custom serving stack. The source package excludes the original compiled CUDA binary and local compiler caches; a clean rebuild must be checked independently.

How the edit works

Each edited MLP down projection receives the same fixed linear correction at every token:

W_new = W - alpha * outer(u, coeff)
layers = [18, 49, 50, 51]
global alpha = 1.5

V1's three corrections are preserved. The new layer-18 selector has 32 supported input columns and was fitted using authored productive-repair versus repetitive-checking prefixes. Its calibration label, 0.5%, refers to productive-prefix correction calibration, not a percentage of model weights or an expected quality gain. No HLE answer labels were used to fit the vectors; HLE development outcomes did influence which candidate was selected.

The evaluated sidecar computes the correction in FP32 and casts it to BF16 before subtraction and tensor-parallel all-reduce. The merged export materializes the algebra in F32 matrices. Different matmul kernels and rounding mean these representations are not asserted to be bitwise or behaviorally equivalent.

Provenance and limits

• Source: HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF, revision 993a5971fda8f30dd1b7eb2654792ba4415c7460.
• Original packed-base SHA256: ba36dc3c2b2ff5e0aa5d71092a8894546996a6a119ae391803dda07cdc08516d.
• V2 delta SHA256: c12820b9bc65e691853a1073c8647fed47cb5521ef3fbcd19817d684006debba.
• Merged GGUF SHA256: 8c046db3518f3711f797a832153362f413aa29e9ab8aef675f88a0842291c1a1; 19,104,353,280 bytes.
• All 64 original blocks execute once. There is no prompt-dependent controller, answer lookup or repeated-block route in this release.
• The name follows the upstream Qwen3.8 label; architecture metadata is qwen35. “4-bit” names a mixed-precision family, and the merged artifact explicitly retains F32 edited matrices.

“Calmer” describes observed text behavior in selected tests, not a measurement of anxiety or wellbeing. Qualitative proof reviews and publication preparation used AI assistance; there was no independent human grading panel. Full gated benchmark prompts, reference answers and raw traces are not redistributed. Public metrics, selected excerpts, failures and reconstruction identifiers are in the research repository.

Credit to Qwen/Alibaba Cloud, HauhauCS and the authors of the cited research. Original project material is Apache-2.0; third-party source retains its own notices. See NOTICE.txt and runtime/NOTICE.txt.