Phil2Sat/Omega-Convergence-27B-v1.0-I-Compact-MXFP4_ALPHA_TESTING

Omega-Convergence-27B-v1.0-SOMPOA-SSMFIX-I-Mini-MXFP4

Experimental Frankenquant — abliterated, uncensored, brain-fixed, block-assembled and RP-optimized.

This is an experimental Frankenquant of ReadyArt/Omega-Convergence-27B-v1.0.

The model is assembled from the Omega base with selected block/tensor transfers from SOMPOA and SSMFIX, followed by a handcrafted RP-focused quantization pass.

The goal is simple:

Keep the RP behavior while pushing the model down to a compact ~4.26 BPW footprint without blindly destroying the tensors that actually matter.

Recommended

Omega-Convergence-27B-v1.0-SOMPOA-SSMFIX-I-Mini-MXFP4.gguf

Latest!!

• Quantization: I-Mini / handcrafted tensor override / MXFP4
• File size: 13,884.57 MiB
• Effective size: 4.26 BPW
• Format: GGUF
• Target: Roleplay

This is not a conventional uniform quantization.

The tensor allocation was deliberately handcrafted around RP calibration statistics, activation behavior, SSM stability, and layer depth.

The idea is to avoid the quality loss and "stupidity" that can appear when simply throwing a standard low-bit quant at the entire model.

---

What is this?

The model is essentially a Frankenstein assembly:

• Omega-Convergence provides the main base/body.
• SOMPOA / Absolute Heresy provides selected transferred attention components.
• SSMFIX provides selected transferred SSM-related components.
• A custom RP calibration pass determines where precision should be spent.
• A handcrafted tensor override assigns different quantization formats to different tensors and layers.
• A fixed Qwen Chat Template 22.1 is included.

The SSM components are not merely a reference or inspiration here. They are part of the actual block/tensor transfer process, just like the SOMPOA components.

---

SSMFIX

Qwen 3.5 / 3.8 hybrid models combine full-attention layers with GatedDeltaNet SSM layers.

The SSM recurrence is governed in part by the 1D convolution weights:

linear_attn.conv1d.weight

The SSMFIX work identified abnormal conv1d statistics in several late Qwen3.8 layers.

Based on the SSMFIX work from redashes/Qwen3.8-27B-BF16-SSMFIX, eight layers showed significantly inflated conv1d standard deviations compared with the approximately 0.042 sibling-layer median.

The corrected values were:

Layer	α applied	Post-fix std
52	0.5901	0.0471
53	0.5548	0.0437
56	0.5449	0.0425
57	0.5357	0.0410
58	0.6097	0.0432
60	0.4814	0.0398
61	0.6533	0.0420
62	0.6186	0.0452

The SSMFIX components are incorporated into this Frankenquant through the same general block/tensor transfer philosophy used for the SOMPOA components.

---

Handcrafted RP Tensor Override

The main difference from a conventional quantization is the handcrafted tensor allocation.

Instead of treating every layer equally, a special RP calibration dataset was used to examine which tensors and layers are activated most strongly.

The resulting statistics were used to manually decide where additional precision should be spent.

One particularly obvious observation was the activation growth in the deeper layers:

Layer	Σ(Act²)	My choice
63	55,505.29	Q5_K ✅
62	12,255.16	Q5_K ✅
61	6,913.21	Q5_K ✅
60	5,122.65	Q5_K ✅
59	5,147.63	Q5_K ✅
58	4,986.85	Q5_K ✅
57	2,862.64	MXFP4 (my practical limit)
52	2,246.74	MXFP4
35	486.03	MXFP4
0	1.21	MXFP4

This is only one simple observation from the analysis. There is considerably more going on across the SSM layers and other tensor groups.

The general strategy is:

The deeper and more activation-heavy a component is, the more aggressively its precision is protected.

The SSM blocks are considered particularly critical, and later blocks therefore receive progressively higher precision.

---

Quantization Strategy

This is intentionally not a uniform "everything Q4" or "everything Q5" quant.

Different tensors use:

• F32
• Q8_0
• Q5_K
• IQ3_S
• MXFP4

depending on their role, activation behavior, layer position, and observed importance.

In particular:

Protected

• Normalization tensors
• Critical SSM parameters
• SSM convolution weights
• Late/deep attention tensors
• High-activation late-layer tensors

Aggressively quantized

• Large FFN tensors where calibration indicated that MXFP4 was acceptable
• Less activation-heavy components
• Selected early/middle-layer tensors

The result is an intentionally asymmetric quantization profile.

---

Actual Build Process

The actual process was:

1. Get the Q8_0 Omega base

Start from the Q8_0 version of the Omega model.

2. Apply the fixed Chat Template

Add Chat Template 22.1 from froggeric/Qwen-Fixed-Chat-Templates.

3. Transfer SOMPOA attention tensors

Transfer:

• attn_output
• ffn_down

from Q8_0 SOMPOA into the Q8_0 Omega model.

SOMPOA source:

MuXodious/Qwen3.8-27B-absolute-heresy-GGUF

4. Transfer SSMFIX components

The SSMFIX components are incorporated through block/tensor transfer as well.

SSMFIX source:

redashes/Qwen3.8-27B-BF16-SSMFIX

5. RP calibration

Use a special RP calibration dataset to obtain imatrix activation statistics.

The purpose is to determine what is actually activated most strongly during RP instead of relying solely on generic calibration assumptions.

6. Craft the tensor override

Use the resulting activation statistics to manually construct tensor_override.txt.

7. Quantize with imatrix

Run llama-quantize with:

• RP calibration imatrix
• MrAdemacher imatrix statistics
• Bartowski imatrix statistics
• 1 calibration chunk
• ctx = 12288

over the resulting Q8_0 model.

8. Final handcrafted quantization

Run the final Q8_0 → mixed-format quantization using the handcrafted tensor_override.txt.

---

Tensor Override

The final handcrafted override:

^token_embd\.weight$=Q5_K
^output\.weight$=Q5_K
^output_norm\.weight$=F32

^blk\..*\.attn_norm\.weight$=F32
^blk\..*\.post_attention_norm\.weight$=F32
^blk\..*\.attn_k_norm\.weight$=F32
^blk\..*\.attn_q_norm\.weight$=F32

^blk\..*\.ssm_a$=F32
^blk\..*\.ssm_alpha\.weight$=Q8_0
^blk\..*\.ssm_beta\.weight$=Q8_0
^blk\..*\.ssm_conv1d\.weight$=F32
^blk\..*\.ssm_dt\.bias$=F32
^blk\..*\.ssm_norm\.weight$=F32

^blk\.(58|59|60|61|62|63)\.ffn_down\.weight$=MXFP4
^blk\.(58|59|60|61|62|63)\.ffn_gate\.weight$=MXFP4
^blk\.(58|59|60|61|62|63)\.ffn_up\.weight$=Q5_K
^blk\.(58|59|60|61|62|63)\.attn_gate\.weight$=Q5_K
^blk\.(58|59|60|61|62|63)\.attn_qkv\.weight$=Q5_K
^blk\.(58|59|60|61|62|63)\.ssm_out\.weight$=Q5_K

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|35|38|39|40|41|42|43|47|51|52|53|54|55|56|57)\.ffn_down\.weight$=IQ3_S
^blk\.(33|34|36|37|44|45|46|48|49|50)\.ffn_down\.weight$=MXFP4

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|35|38|39|40|41|42|43|47|51|52|53|54|55|56|57)\.ffn_gate\.weight$=IQ3_S
^blk\.(33|34|36|37|44|45|46|48|49|50)\.ffn_gate\.weight$=MXFP4

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36|37|38|39|40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57)\.ffn_up\.weight$=MXFP4

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36|37|38|39|40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57)\.attn_qkv\.weight$=MXFP4

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|35|38|39|40|41|42|43|47|51|52|53|54|55|56|57)\.attn_gate\.weight$=MXFP4

^blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35|36|37|38|39|40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57)\.ssm_out\.weight$=MXFP4

^blk\.(33|34|36|37|44|45|46|48|49|50)\.attn_gate\.weight$=Q5_K

^blk\.(3|7|11|15|19|23|27|31|35|39|43|47|51|55|59|63)\.attn_k\.weight$=Q5_K
^blk\.(3|7|11|15|19|23|27|31|35|39|43|47|51|55|59|63)\.attn_v\.weight$=Q5_K
^blk\.(3|7|11|15|19|23|27|31|35|39|43|47|51|55|59|63)\.attn_q\.weight$=Q5_K
^blk\.(3|7|11|15|19|23|27|31|35|39|43|47|51|55|59|63)\.attn_output\.weight$=Q5_K

^blk\.64\.attn_k\.weight$=MXFP4
^blk\.64\.attn_k_norm\.weight$=F32
^blk\.64\.attn_norm\.weight$=F32
^blk\.64\.attn_q\.weight$=MXFP4
^blk\.64\.attn_q_norm\.weight$=F32
^blk\.64\.attn_v\.weight$=MXFP4
^blk\.64\.ffn_down\.weight$=MXFP4
^blk\.64\.ffn_gate\.weight$=MXFP4
^blk\.64\.ffn_up\.weight$=MXFP4
^blk\.64\.nextn\.eh_proj\.weight$=Q8_0
^blk\.64\.nextn\.enorm\.weight$=F32
^blk\.64\.nextn\.hnorm\.weight$=F32
^blk\.64\.nextn\.shared_head_norm\.weight$=F32
^blk\.64\.post_attention_norm\.weight$=F32
^blk\.64\.attn_output\.weight$=MXFP4

---

Why the strange quantization?

Because the model is not equally sensitive everywhere.

A standard quantizer has to make compromises across the entire model. This build instead spends precision where the calibration data suggests that it matters.

The particularly important observations were:

• SSM tensors are sensitive.
• Late layers show substantially higher activation magnitudes.
• Some attention tensors deserve more protection than their raw parameter count would suggest.
• MXFP4 is surprisingly effective for many large tensors when applied selectively.
• Uniform low-bit quantization can throw away useful behavior unnecessarily.
• A small number of Q5_K/Q8_0/F32 tensors can be used to protect critical components while keeping the overall model around 4.26 BPW.

So the philosophy is basically:

Don't waste bits on tensors that don't need them. Don't save bits on tensors that do.

---

Credits & Acknowledgements

This Frankenquant builds directly on the work of several other people.

Omega / Base

Huge thanks to FrenzyBiscuit and Ready.Art for Omega-Convergence-27B-v1.0.

Omega is the main base/body of this Frankenquant.

The ReadyArt ecosystem also provides the broader Omega work this build is based on. FrenzyBiscuit is credited as a contributor/creator in the ReadyArt Omega project.

SOMPOA

Huge thanks to MuXodious for Qwen3.8-27B-absolute-heresy-GGUF.

SOMPOA is one of the actual block/tensor sources used in this build.

Selected attention components are transferred from the SOMPOA-derived Q8_0 model into the Omega Q8_0 base before the final quantization.

SSMFIX

Huge thanks to redashes for Qwen3.8-27B-BF16-SSMFIX.

SSMFIX is likewise an actual block/tensor source for this Frankenquant.

The SSMFIX work was especially important because it exposed the abnormal SSM conv1d behavior that needed to be investigated before aggressively quantizing these components.

LuffyTheFox

A special thanks to LuffyTheFox.

The investigation into the Qwen3.8 SSM architecture started because of LuffyTheFox's Qwen3.8 discussion post, which first pointed me toward the SSM issue.

That led to the deeper investigation of:

• GatedDeltaNet / SSM layers
• linear_attn.conv1d.weight
• abnormal standard deviations
• activation behavior
• late-layer sensitivity
• and ultimately the decision to protect SSM components much more aggressively during quantization.

So while LuffyTheFox's work is not directly a block source in this model, it was the starting point that led me down the SSM rabbit hole.

Chat Template

Thanks to froggeric for the fixed Qwen templates:

Qwen-Fixed-Chat-Templates

This build uses Chat Template 22.1.

---

Credits in Short

FrenzyBiscuit / Ready.Art → Omega base / main model foundation MuXodious / SOMPOA → transferred attention components redashes / SSMFIX → transferred SSM components and SSM correction work LuffyTheFox → led me to investigate the Qwen3.8 SSM issue froggeric → fixed Qwen Chat Template 22.1 Me → block assembly, tensor transfers, RP calibration, activation analysis, handcrafted tensor allocation, quantization strategy and final Frankenquant

Huge thanks to everyone involved.

This model is very much a "standing on the shoulders of other people's experiments" kind of build.

---

Disclaimer

This is an experimental Frankenquant.

The quantization profile is intentionally unusual and handcrafted. The allocation decisions are based on RP calibration, activation statistics, SSM observations, and practical testing rather than a formal benchmark proving that every selected tensor format is optimal.

If you find something broken, stupid, unstable, or unexpectedly brilliant:

that's probably part of the experiment.