All my model releases so far, in one place

Thu, 08 Oct 2026 16:49:00 +0000

I’ve been posting model release updates as they happened. This collects everything through September in one place, organized by model family. From here on, I’ll post them as they go.

Qwen3.6-27B Fable-Fusion NVFP4

Four tiers: LOW, MEDIUM, HIGH, VERY-HIGH. This is a splice of maci0’s native NVFP4 backbone (all 64 transformer blocks, no re-quantization round trip) and DavidAU’s low-memory heads, with the MTP speculative head embedded. Model credits: Alibaba Qwen, DavidAU, maci0. My part is the splice. ~640 t/s prefill, ~23 t/s decode on 2x16GB Blackwell. Apache-2.0.

Link:

https://www.huggingface.co/esatapedico/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP-NVFP4-GGUF

Qwen3.8-27B NVFP4-MTP family

Seven tiers from ORIG (the source-preserving 33 GB conversion) down to compact files around 15 GB, all sharing one byte-identical NVFP4 backbone. The MTP speculative head is embedded in every file, so no separate drafter is needed. Native vision-language projector included. Upstream: Unsloth’s NVFP4 quantization of Alibaba’s Qwen3.8-27B; my part is conversion and tiering. Every tier fits 2x16GB Blackwell.

Additional tiers added later:

BUDGET and STARVED: the two smallest ways to run this on a single 16 GB Blackwell card. Speculative head stripped, extras trimmed to minimum. ~2288 t/s prefill, 27 t/s decode, peaking at 15.9 GiB on an RTX 5070 Ti. Vision works too.
VERY-LOW: smallest compact tier (~14.9 GB), everything trimmed for a tight 16 GB card.
LOW-MINUS: between LOW and VERY-LOW. I bumped the tensors that matter most (leaving MTP head lowest to favor quality over squeezed performance).
MID-HIGH: Q8 for lm_head, token_embd, and MTP head, attention stays NVFP4. Lighter than HIGH+ tiers with comparable decode speed and the highest MTP accuracy we observed (preliminary, not a proper benchmark).
HIGHEST: closest to source precision, Q8 attention on native NVFP4 backbone with BF16 embeddings (~23 GB). At 32k context on a single 16 GB card: 1250 t/s prefill, 23 t/s decode, comfortable up to 131k context.

Links:

Independent benchmarks: PierPaolo Pernici ran the tiers on an RTX 5090 (32 GB), 192K context, q8_0 KV. LOW hit ~148 avg tok/s (203 peak), MEDIUM ~134, VERY-HIGH ~112. MTP drafting sweet spot at n-max 4. 70k-token long-context: 109.7 tok/s on LOW. Full MMLU: 82.41%. His comparison against other NVFP4-MTP conversions on the Hub found the best results from our files, especially with an alternative MTP head attached (cross-file Q5_K draft, 112.6 tok/s on long context).

Link:

https://gist.github.com/PierpaoloPernici/2e6f6f42965d531b364f39ab0e8a52ad

Qwen3.8-27B NVFP4-MTP with SSMFIX

Eight NVFP4 GGUF tiers with the community fix people call “SSMFIX” applied. Eight small attention kernels get scaled down by per-layer factors. The idea came from LuffyTheFox’s analysis, turned into a recipe and BF16 checkpoint by redashes, with FGDumitru landing on the same layers independently. All credit for the idea goes to them. I only made the NVFP4 GGUF tiers.

This is unverified. Some users report it helps, some careful benchmarks show it making things worse. My own single runs were genuinely mixed. Good results are not guaranteed, and worse results are not unlikely.

Link:

https://www.huggingface.co/esatapedico/Qwen3.8-27B-NVFP4-MTP-SSMFIX-GGUF

Muse-Glimmer-30B NVFP4 GGUF

Six tiers, VERY-LOW to HIGHEST, with a native NVFP4 backbone. All 416 attention/FFN weights kept as native NVFP4 (GGML type 40), only lm_head and token_embd tiered. Upstream: Meta’s Muse Glimmer 30B (52 layers, sliding window, 131K context) and Red Hat AI’s NVFP4 quantization.

The NVFP4 conversion needed a Q/K layout fix that isn’t merged upstream yet; gabrielcosi’s llama.cpp fork is the reliable runtime for now (issue #27178).

At short context on 2x16GB Blackwell: ~2500-2600 t/s prefill, ~28-33 t/s generation. The DFlash drafter gives ~1.5x faster decode. My part: conversion, QK-norm absorption, tiering, verification. Apache-2.0.

Link:

https://www.huggingface.co/esatapedico/Muse-Glimmer-30B-NVFP4-GGUF

Qwen3.8-27B Cold Fusion GAIN V1.1 NVFP4

Six tiers with a native NVFP4 backbone. Upstream: DavidAU’s Cold Fusion tune (GAIN + Unsloth training) and windowsxp811203’s NVFP4 checkpoint. All credit for the tune goes to them. My part is the GGUF side: one byte-stable 448-tensor NVFP4 backbone across all tiers, from VERY-LOW at 14.9 GB up to VERY-HIGH with BF16 heads, MTP head baked into every file. Currently experimenting with MID-HIGH as daily driver.

Link:

https://www.huggingface.co/esatapedico/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NVFP4-GGUF

Qwen3.8-27B TURBO 735-882 NVFP4 GGUF

DavidAU’s TURBO 735-882 tune, converted to NVFP4 GGUF in eight tiers, 15 to 21 GB. One shared NVFP4 backbone, MTP head baked into every file. Upstream: Alibaba Qwen, DavidAU tune. My part is conversion and tiering only. Quick local checks look flat across the ladder. Apache-2.0.

Also available: budget files without MTP for single 16 GB cards (14.6 to 14.7 GB), plus the NVFP4 checkpoint the GGUFs were built from.

Links:

That’s the whole backlog. So far.

esatapedico's laboratory

All my model releases so far, in one place

Qwen3.6-27B Fable-Fusion NVFP4

Qwen3.8-27B NVFP4-MTP family

Qwen3.8-27B NVFP4-MTP with SSMFIX

Muse-Glimmer-30B NVFP4 GGUF

Qwen3.8-27B Cold Fusion GAIN V1.1 NVFP4

Qwen3.8-27B TURBO 735-882 NVFP4 GGUF