Calmer Qwen3.8 Flash Next ยท Gyro-M

Two fixed weight edits at layers 44 and 45. Selected for daily use after a four-question production-runtime comparison used 35.3% fewer generated tokens, with all four final answers correct.

Research and experiment archive ยท Earlier Calmer Qwen research ยท Calmer V2 27B

This experimental derivative by Xananthium edits the Q8_0 shared-expert down projections in the AgentionAI Gyro-M quantization of Qwen3.8-Flash-Next. All layer indices are zero-based. Layers 2 and 46 are original in this release.

The measurements below used this materialized GGUF in Strata on two RTX 3090s, with medium reasoning, temperature 1.3, and the existing separate runtime control vector enabled in both arms. That vector is not baked into the published weight edits. These are exploratory results on reused questions, not full benchmark scores or an evaluation without runtime steering.

What to download

Artifact Purpose
Calmer-Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf Complete model with the selected 44+45 edits; 91,965,774,272 bytes
release_manifest.json Pinned upstream identity, tensor offsets, original and edited hashes, formulas, and evaluation scope
patches/ Exact original and edited Q8 tensor payloads, plus the saved layer-45 direction
evaluation_summary.json Completed main-panel comparisons, loop probes, and interrupted variants
tested_runtime.json Recorded production settings, required separate assets, and natural-completion settings
LICENSE, NOTICE.txt Qwen Community License 1.0 and attribution

The filename's TQ2_0 label identifies the upstream Gyro-M size class. This is a mixed-precision custom GGUF, not uniformly two-bit weights. The two edited matrices remain Q8_0. They were decoded, edited, and encoded back into their existing representation; the rest of the model was not converted to eight-bit.

Observed results

The matched panel contains two difficult HLE questions and two GSM8K arithmetic controls, at seed 20261118. Generated-token counts include reasoning and the final answer.

Case Unedited weights Calmer 44+45
Economics 21,315 19,756
Mathematics 53,084 27,789
Hotel arithmetic 581 657
Pens arithmetic 555 639
Total 75,535 48,841 (โˆ’35.3%)
Correct final answers 4/4 4/4
Natural completions 4/4 4/4

The math response used 47.7% fewer tokens and economics 7.3% fewer. The two arithmetic controls together used 14.1% more, so shortening was not universal. Raw occurrences of โ€œwaitโ€ fell from 115 to 54 across these four outputs; that count is descriptive, not a diagnosis or an independent measure of reasoning quality.

Eight short constructed continuation/tool-correction probes scored 7/8 โ†’ 8/8 under the saved automatic grader. In the Git probe, the original repeated git diff after explicit wrong-directory feedback; 44+45 emitted the requested git -C /app/project diff. These were simulated histories: no generated shell commands were executed.

The archive includes 45-only, 45+46, 44+45, 44+45+46, full layer-2 +44+45, and partial layer-2/layer-46 variants, along with the unedited baseline. Two longer math runs were manually interrupted above 70,000 observed tokens and remain marked incomplete. Earlier 8,192/32,768-token studies had different forced-ending settings and are reported separately.

How the edit works

Let d45 be the leading left singular vector of the original decoded layer-45 shared-expert down-projection matrix. Apply:

W45_new = W45 - outer(d45, d45.T @ W45)
W44_new = W44 - outer(d45, d45.T @ W44)

Both strengths are 1.0. Layer 44 uses the original layer-45 direction, not its own leading singular vector. The selected tensors are blk.44.ffn_down_shexp.weight and blk.45.ffn_down_shexp.weight, each shaped 2560 ร— 640 and encoded as Q8_0. Quantization rounding leaves a small residual in the removed direction. There is no prompt-dependent edit, answer lookup, or learned stopping controller.

The direction came from static weight SVD, not a fitted hesitation detector. Development outcomes influenced candidate selection. The experiments do not establish that a large singular direction is a general repetition feature.

Download and use

hf download Xananthium/Calmer-Qwen3.8-Flash-Next \
  --local-dir Calmer-Qwen3.8-Flash-Next

Use an engine supporting the upstream custom Gyro tensor types and qwen4exp architecture. See the source quantization's runtime instructions. Stock GGUF support alone does not establish compatibility. The measured engine identified itself as Strata 0.1.59.1; this is an engine version, not the desktop application's version.

The production configuration used 262,144 context, Q4_0 KV, two GPUs split at layer 24, four parallel slots, four-token MTP speculation, and the existing expert pack. Its separate Huihui control vector was projected per layer over layers 4โ€“44 at scale 1.0. The archive records the asset identities and configuration. The runtime vector is separately available from alesha-pro.

Set temperature 1.3, top-p 0.95, top-k 20, min-p 0, presence penalty 0, repetition penalty 1, and medium reasoning. The natural-completion panel disabled answer-room reservation, reasoning-budget closure, close retry, loop recovery, and repeated-token stopping. Its 75,000 output-token ceiling was per test request, not a short thinking budget or a global production cap.

When using the recorded Strata native/PLE layout, point both --native and --ple-gguf at the same edited GGUF. A mismatched original/edited pair caused a duplicate split-shard startup error during the first experiment. The saved historical scripts are an experiment record; use the portable reconstruction tool in the research repository for a new model copy.

Provenance and limits

  • Quantized source: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF, revision 8efdc41d092da63763f39173b0fff80eac1c0898.
  • Source file: Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf, SHA256 7aa36d15131d255179ab6c87fa9c9f81a0b541fc9e81e959d4c3e29eb20feb5f.
  • Official base: Qwen/Qwen3.8-Flash-Next, revision 812d96e10a0dd82b69d749e68693698cf6c36daa.
  • Released GGUF SHA256: 60f5dd9fb65eb9997de6b3879a23482615d8a51689f091de0f4b8e9a749dfe16.
  • Only layers 44 and 45 are selected. The released payload hashes match the restored production tensors; layers 2 and 46 match their originals.
  • This is the materialized Q8 edit representation used in the reported comparison, unlike the earlier 27B Calmer sidecar experiments.

The panel is small, adaptively explored, and largely one seed. Concurrent production traffic and speculative decoding can affect repeatability. Correct final choices are not a proof of every step in a long explanation. Vision was not evaluated. โ€œCalmerโ€ names observed text behavior in these tests; it does not measure anxiety or wellbeing.

Credit to Qwen/Alibaba Cloud for the base model, AgentionAI for the quantization and compatible runtime, and the HLE/GSM8K authors for the benchmark examples. Publication, tooling, and qualitative review used AI assistance. This is an experimental derivative, not an official Qwen release. Weights and derived tensor payloads retain the Qwen Community License 1.0; see LICENSE.

The public research export follows the earlier Calmer archive: HLE prompts, keys, and full traces remain local under the dataset authorsโ€™ requested convention. Every runโ€™s measurements and hashes, the other raw outputs, and unsuccessful variants are included.

Downloads last month
204
GGUF
Model size
177B params
Architecture
qwen4exp
Hardware compatibility
Log In to add your hardware

2-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Xananthium/Calmer-Qwen3.8-Flash-Next

Quantized
(2)
this model