Compare model versions

Model: peculiar-ragdoll/Tiel-Coder-35B-A3B-MLX-oQ4e

Earlier snapshot Snapshot #6047 Repository revision 058d0c1 (opens in a new tab) Recorded at 30 Sep 2026, 07:04 UTC Unchanged through 29 Oct 2026, 17:01 UTC

Later snapshot Snapshot #6586 Repository revision 24ee72e (opens in a new tab) Recorded at 30 Oct 2026, 09:36 UTC Unchanged through 04 Nov 2026, 10:46 UTC

Summary

Changed: 5 weight files changed
Changed: 638 configuration values changed
Changed: Model card modified (2 sections)
Changed: Repository revision 058d0c1 → 24ee72e
Changed: 2 other files changed
Unchanged: Licence unchanged
Unchanged: Remote code unchanged
Unchanged: Access unchanged

Licence

Unchanged

No change

MIT MIT

PinDiff reports the declared licence and licence files. It does not interpret licence terms.

Weights

Changed

5 files changed

5 changed hash

Weight files
File	Status	Before	After
`model-00001-of-00005.safetensors`	CHANGED	SHA256 (LFS) `a85cfa9cf3a90254e3ad129543d3b0598d7c32bc23f5863afc11639508982b63` 5.10 GB	SHA256 (LFS) `b25dd783d843b29d252baae57adf4b41280f5c12f197a3cd6d522be617efe413` 5.06 GB (−36.73 MB)
`model-00002-of-00005.safetensors`	CHANGED	SHA256 (LFS) `ccf6b67d3ddbac7cbe479fdb3d393340397e081f14112b5401b8c15e2f038f7a` 5.13 GB	SHA256 (LFS) `d4d1d8873569f072f4b9829ca87521d41a5e4d3eb556bc5088652a02e781ce17` 5.13 GB
`model-00003-of-00005.safetensors`	CHANGED	SHA256 (LFS) `9679f07dfdfc7497e184ee6059e360700905cbbec0801f08055402233a46b0cf` 5.13 GB	SHA256 (LFS) `2cc037cdc77468f08d1ca4a06e519de18fa0ab450b46d1f13d62f57c4173e821` 5.13 GB
`model-00004-of-00005.safetensors`	CHANGED	SHA256 (LFS) `b8ef5e73275fd0cef1eb14948198125e66969e73e6a6cdcf9ad9abebfc252f29` 5.13 GB	SHA256 (LFS) `f855fcaba0521326f869ae4217bbece3ea4a9c698f303c5c3ba177d6fcc0e436` 5.13 GB
`model-00005-of-00005.safetensors`	CHANGED	SHA256 (LFS) `822736cf4e61adefef32f83f0c9a7a15d0ba8a55c4c51182c093eaebaf96ff34` 603.98 MB	SHA256 (LFS) `a3ea673928bce7c61a97d6242b67f25cd59f4111ce9285bff8cc9021918e7fee` 603.98 MB

Remote code / configuration

Changed

638 configuration values changed

Remote code No No

`config.json` Changed

Removed: - quantization.language_model.lm_head.bits: 8

Added: + quantization.language_model.lm_head.bits: 6

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_a.bits: 8

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_b.bits: 8

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_qkv.bits: 6

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_qkv.group_size: 128

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_qkv.group_size: 64

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.0.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.0.linear_attn.in_proj_z.group_size: 64

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.1.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.1.linear_attn.in_proj_z.group_size: 64

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_a.bits: 8

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_b.bits: 8

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_qkv.bits: 6

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_qkv.group_size: 128

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_qkv.group_size: 64

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.10.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.10.linear_attn.in_proj_z.group_size: 64

Added: + quantization.language_model.model.layers.11.self_attn.k_proj.bits: 5

Added: + quantization.language_model.model.layers.11.self_attn.k_proj.group_size: 64

Added: + quantization.language_model.model.layers.11.self_attn.k_proj.mode: "affine"

Added: + quantization.language_model.model.layers.11.self_attn.o_proj.bits: 5

Added: + quantization.language_model.model.layers.11.self_attn.o_proj.group_size: 64

Added: + quantization.language_model.model.layers.11.self_attn.o_proj.mode: "affine"

Added: + quantization.language_model.model.layers.11.self_attn.q_proj.bits: 5

Added: + quantization.language_model.model.layers.11.self_attn.q_proj.group_size: 64

Added: + quantization.language_model.model.layers.11.self_attn.q_proj.mode: "affine"

Added: + quantization.language_model.model.layers.11.self_attn.v_proj.bits: 5

Added: + quantization.language_model.model.layers.11.self_attn.v_proj.group_size: 64

Added: + quantization.language_model.model.layers.11.self_attn.v_proj.mode: "affine"

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.12.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.12.linear_attn.in_proj_z.group_size: 64

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.14.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.14.linear_attn.in_proj_z.group_size: 64

Added: + quantization.language_model.model.layers.15.self_attn.k_proj.bits: 5

Added: + quantization.language_model.model.layers.15.self_attn.k_proj.group_size: 64

Added: + quantization.language_model.model.layers.15.self_attn.k_proj.mode: "affine"

Added: + quantization.language_model.model.layers.15.self_attn.o_proj.bits: 5

Added: + quantization.language_model.model.layers.15.self_attn.o_proj.group_size: 64

Added: + quantization.language_model.model.layers.15.self_attn.o_proj.mode: "affine"

Added: + quantization.language_model.model.layers.15.self_attn.q_proj.bits: 5

Added: + quantization.language_model.model.layers.15.self_attn.q_proj.group_size: 64

Added: + quantization.language_model.model.layers.15.self_attn.q_proj.mode: "affine"

Added: + quantization.language_model.model.layers.15.self_attn.v_proj.bits: 5

Added: + quantization.language_model.model.layers.15.self_attn.v_proj.group_size: 64

Added: + quantization.language_model.model.layers.15.self_attn.v_proj.mode: "affine"

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.16.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.16.linear_attn.in_proj_z.group_size: 64

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.17.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.17.linear_attn.in_proj_z.group_size: 64

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_a.bits: 5

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_a.bits: 6

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_a.group_size: 128

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_a.group_size: 64

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_b.bits: 5

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_b.bits: 6

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_b.group_size: 128

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_b.group_size: 64

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_qkv.bits: 5

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_qkv.group_size: 64

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_qkv.mode: "affine"

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_z.bits: 5

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_z.bits: 6

Removed: - quantization.language_model.model.layers.18.linear_attn.in_proj_z.group_size: 128

Added: + quantization.language_model.model.layers.18.linear_attn.in_proj_z.group_size: 64

Added: + quantization.language_model.model.layers.19.self_attn.k_proj.bits: 5

Added: + quantization.language_model.model.layers.19.self_attn.k_proj.group_size: 64

Added: + quantization.language_model.model.layers.19.self_attn.k_proj.mode: "affine"

Added: + quantization.language_model.model.layers.19.self_attn.o_proj.bits: 5

Added: + quantization.language_model.model.layers.19.self_attn.o_proj.group_size: 64

Added: + quantization.language_model.model.layers.19.self_attn.o_proj.mode: "affine"

Added: + quantization.language_model.model.layers.19.self_attn.q_proj.bits: 5

Added: + quantization.language_model.model.layers.19.self_attn.q_proj.group_size: 64

Added: + quantization.language_model.model.layers.19.self_attn.q_proj.mode: "affine"

+538 more changes not shown.

`generation_config.json` Unchanged

Access / gating

Unchanged

No change

Public Public

Model card

Changed

Model card modified (2 sections)

No multi-token-prediction head Changed +6 −3

Diff

Removed: - Ornith-1.5's GGUF conversion carries an MTP (`nextn`) block whose weights are untrained — a fresh

Removed: - random initialization, which we measured and removed from the GGUF ladder. It is not in the

Removed: - safetensors this build quantizes from, so there was nothing to strip here.

Added: + This build has no MTP head. The safetensors it quantizes from carry Ornith-1.5's trained `nextn`

Added: + head, but this plain build leaves it out; the sibling

Added: + [oQ4e-MTP build](https://huggingface.co/peculiar-ragdoll/Tiel-Coder-35B-A3B-MLX-oQ4e-MTP) keeps it, for

Added: + runtimes that can draft with it. (Ornith's *original* GGUF conversion once shipped an untrained,

Added: + randomly-initialized `nextn` block, which we measured and removed from the earlier GGUF ladder; Ornith

Added: + later published the trained head.)

Before

Ornith-1.5's GGUF conversion carries an MTP (`nextn`) block whose weights are untrained — a fresh
random initialization, which we measured and removed from the GGUF ladder. It is not in the
safetensors this build quantizes from, so there was nothing to strip here.

After

This build has no MTP head. The safetensors it quantizes from carry Ornith-1.5's trained `nextn`
head, but this plain build leaves it out; the sibling
[oQ4e-MTP build](https://huggingface.co/peculiar-ragdoll/Tiel-Coder-35B-A3B-MLX-oQ4e-MTP) keeps it, for
runtimes that can draft with it. (Ornith's *original* GGUF conversion once shipped an untrained,
randomly-initialized `nextn` block, which we measured and removed from the earlier GGUF ladder; Ornith
later published the trained head.)

How the quantization was done Changed +4 −1

Diff

⋯ 1 unchanged line

non-quantization, plus an imatrix pass — the "e" — that measures which weights carry the most

signal before deciding what to keep at higher precision. It is the same idea as the GGUF ladder's

Unsloth-Dynamic-plus-imatrix recipe, implemented for MLX, but it is *not* the same computation:

Removed: - oQ derives its own importance data rather than consuming the GGUF imatrix we baked.

Added: + oQ derives its own importance data rather than consuming the GGUF imatrix we baked. It derives it from

Added: + the same data, though: the imatrix pass reads exactly the 3,000 chunks of 512 tokens the GGUF imatrix was

Added: + computed on — the start of the TielCoder calibration corpus, in the same order, checked token for token

Added: + against llama.cpp's tokenizer — rather than oMLX's built-in calibration set.

That is the reason for the caveat at the top. Two quantizers pursuing the same goal by different

routes do not land in the same place, and only the GGUF route has been benchmarked.

Before

oQ4e is oMLX's dynamic quantizer: 4-bit base with mixed precision by layer position and selective
non-quantization, plus an imatrix pass — the "e" — that measures which weights carry the most
signal before deciding what to keep at higher precision. It is the same idea as the GGUF ladder's
Unsloth-Dynamic-plus-imatrix recipe, implemented for MLX, but it is *not* the same computation:
oQ derives its own importance data rather than consuming the GGUF imatrix we baked.

That is the reason for the caveat at the top. Two quantizers pursuing the same goal by different
routes do not land in the same place, and only the GGUF route has been benchmarked.

After

oQ4e is oMLX's dynamic quantizer: 4-bit base with mixed precision by layer position and selective
non-quantization, plus an imatrix pass — the "e" — that measures which weights carry the most
signal before deciding what to keep at higher precision. It is the same idea as the GGUF ladder's
Unsloth-Dynamic-plus-imatrix recipe, implemented for MLX, but it is *not* the same computation:
oQ derives its own importance data rather than consuming the GGUF imatrix we baked. It derives it from
the same data, though: the imatrix pass reads exactly the 3,000 chunks of 512 tokens the GGUF imatrix was
computed on — the start of the TielCoder calibration corpus, in the same order, checked token for token
against llama.cpp's tokenizer — rather than oMLX's built-in calibration set.

That is the reason for the caveat at the top. Two quantizers pursuing the same goal by different
routes do not land in the same place, and only the GGUF route has been benchmarked.

7 other sections unchanged.

Repository status

Changed

Repository revision moved

Earlier: Revision 058d0c1 (opens in a new tab) · 22 files · 21.13 GB
Later: Revision 24ee72e (opens in a new tab) · 22 files · 21.09 GB

A new revision does not by itself mean files changed; the categories above show what differs.

Files

Changed

2 other files changed

2 changed hash

Other files
File	Status	Before	After
`model.safetensors.index.json`	CHANGED	Git blob ID `f493ea8d8d06309c9060ca9073b2ed576bb24c28` 200.2 KB	Git blob ID `50ead79b18653f21f8ee502cd83d1ccd30402131` 200.2 KB
`oq_imatrix_report.json`	CHANGED	Git blob ID `62568f0473c1beeb2ee842caba8832af9aa89fc8` 32.1 KB	Git blob ID `f636c3112ca26769fa7345c0dd283d350ca38bd2` 32.1 KB (−16 B)

Evidence observed by PinDiff from the Hugging Face API and stored when recorded. PinDiff compares stored metadata, hashes and text files; it does not download weight files. Change events recorded between these snapshots

Summary

config.json Changed

generation_config.json Unchanged

Diff

Before

After

Diff

Before

After

`config.json` Changed

`generation_config.json` Unchanged