Not a programmer but want to help -- Currently use a Mac Studio M2 Ultra w/192GB Unified Memory

[splatstx]

I have really gotten the AI fever and want to find ways to help, but I can't program like you guys. I have been working on the oQ Quantization and posting them in HF

If there is testing, Quantizations or something else I can help with please let me know. This is probably the best LLM application I have used so far.

[splatstx]

Uploading oQ+ Quantization models here for anyone interested:
https://huggingface.co/splats/models

[HLGRL]

Hello @splatstx

maybe you could upload as oQ6 quant this model?
lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
This model looks promising, thanks in advance for uploading oQ6 quant if you could 👍

[splatstx]

Hello @splatstx

maybe you could upload as oQ6 quant this model? lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled This model looks promising, thanks in advance for uploading oQ6 quant if you could 👍

@HLGRL Let me know if this works for you:

Lordx64 Qwen3.6-35B Claude-Distilled

Build Strategy: Reference-Tethered (Proxy Sensitivity: oQ8e Master)

oQ Level	bpw	Size	Technical Note
oQ8e	8.00	32.5 GB	The precision anchor for the suite.
oQ6e	6.60	28.1 GB	0.0000 drift. 162 logic-locked boosts.
oQ5e	5.67	24.0 GB	0.0000 drift. 352 logic-locked boosts.
oQ4e	4.70	19.7 GB	0.0000 drift. 318 logic-locked boosts (6b/5b staircase).

Project Audit:

• Zero-Drift Architecture: All enhanced tiers (4e, 5e, 6e) were measured against the oQ8 Master. The 0.0000 variance across all levels confirms a near-lossless transition for reasoning logic.
• MoE Stability: Gate tensors and Shared Experts were prioritized for 8-bit pinning to ensure Claude-4.7 reasoning traces remain coherent.
• Context Integrity: Full mlx-vlm sanitize chain applied; conv1d layers preserved in FP16 for 64k window stability.

[HLGRL]

@splatstx
Thank you very much, btw these distilled models using Claude Opus 4.7 reasoning are really night and day vs original Qwen 3.6-35B. I tested oQ6e and it refactored dozens of files without error, where the same task when given to standard Qwen 3.6-35B either goes into infite loop or fail tool calls in Claude Code.
I highly recommend these distilled models and thank you again very much for these quants 👍

[nickcolea]

@splatstx Really great if we could have a Qwen3.6-27B in oQ3.5(e) as the oQ4's don't fit on 24GB Macs.
These look great but are GGUFs: Unsloth UD

Just saw them on HF. THANK YOU! 🤗