What it is
Table with columns: Property, Value| Property | Value |
|---|
| Method | ties (trim, elect sign, disjoint merge) |
| Weights | [0.5, 0.5] — parents A and R equally |
| Density | 0.2 |
| Parents | Qwen/Qwen3.5-9B (A) + xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic (R) |
| Architecture | Qwen3_5ForConditionalGeneration, hidden 4096, 32 layers |
| Tensors | 775, all merged and verified (dropped_tensors = {}) |
| MTP | 15 mtp.* tensors merged via TIES from both parents |
| Shards | 4 safetensors, 17.98 GiB total |
| Precision | BF16 throughout |
| Tensor-set mode | strict — both parents carry the same 775 names |
| Licence | apache-2.0 |
Shard contents are 42 / 168 / 160 / 405 tensors and match
model.safetensors.index.json.
Parents
- A —
Qwen/Qwen3.5-9B @ c20223623576 — 4 shards, 775 tensors, 15 mtp.*,
17.98 GiB, apache-2.0. The anchor. Vendor benchmarks: GPQA 81.7,
LiveCodeBench v6 65.6, MMLU-Pro 82.5.
- R —
xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic @
fbccdc6adf70 — 5 shards, 775 tensors, 15 mtp.*, 17.98 GiB, apache-2.0.
Reported: GSM8K 94.5, MMLU-Pro avg 90.53.
The benchmark numbers above belong to the parents. They say nothing about
this merge, and must not be read as this model's scores.
This is not KWYN-9B-v0.1
KWYN-9B-v0.1 is reserved for the final trained model — the SFT output built on
a merge. This repository is an intermediate merge one step upstream of that. The
current repository name is KWYN-9B-v0.1; it is queued for rename to
KWYN-9B-v0.1-m1 by the artifact owner. If you are reading a repository still
named KWYN-9B-v0.1 that contains no trained adapter, it is this merge, not the
release.
Tokenizer
The tokenizer is inherited from parent A, vocab_size 248,070.
Parent R contributes 7 tail-only special tokens at ids 248,070–248,076 —
<|audio_start|>, <|audio_end|>, <tts_pad>, <tts_text_bos>,
<tts_text_eod>, <tts_text_bos_single>, <|audio_pad|> — and these were not
written to this artifact. Those ids are unassigned here. Anything that depends
on the audio/TTS special tokens will not work on -m1, and token ids at or above
248,070 should not be assumed to mean what they mean in R.
Intended use
- As the base for behavioural SFT (step P4 of the merge plan), after the merge
passes its stage evaluation.
- As a comparison point while evaluating later merges (
-m2, -m3, tree
variants).
Limits
- Not evaluated. Step P3 of the plan has not run against this artifact. There
are no HumanEval, MBPP, GSM8K, GPQA, planning or tool-use results for
-m1.
Any claim that it improves on stock Qwen/Qwen3.5-9B is unmeasured.
- Not instruction-tuned by us. It is a weight average of two existing
checkpoints; no new data, no preference optimisation, no adapter.
- Inherits both parents' limitations — including R's uncensored behaviour.
R is an "uncensored" derivative, so safety behaviour is not the same as stock
Qwen's and should be re-tested rather than assumed.
- MTP merged, not carried. The 15
mtp.* tensors were TIES-merged like every
other tensor, because both parents carry them. (A future m2 merge will carry
mtp.* verbatim from the anchor instead; see full-plan.md.)
- No quantised builds are published from
-m1 yet.
Licence and attribution
apache-2.0, inherited from the parents.
- Base model and tokenizer: Qwen —
Qwen/Qwen3.5-9B @ c20223623576.
- Second parent:
xtools-at —
xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic @ fbccdc6adf70.
Both parents are apache-2.0. This merge adds no new weights of its own beyond the
TIES combination of theirs.
Provenance
Merge parameters and per-shard verification are recorded in
m1_merge_report.json in this repository: method, weights, density, parent
revisions, tensor_set, dropped_tensors, tensor_count = 775,
verified_tensor_count = 775, shard→tensor mapping, elapsed 3,374 s, peak RSS
2.x GiB (streamed ranged-read path).
Plan of record for the merge track: full-plan.md at the root of the project
repository.