Models

Open-weight models

30 models and 10 datasets on Hugging Face, downloaded more than 16M times. The main language models ship with GGUF builds for llama.cpp, Ollama and LM Studio on the day they come out, and PiVo runs on any CPU.

16Mdownloads on Hugging Face
4.2Mdownloads in the last 30 days
30open-weight models
6,828likes

Hugging Face, November 2026. Downloads count the original weights and the GGUF builds together.

Quark

New

A small hybrid language model: three local layers under each global attention layer, trained on 500B tokens and extended to 128K context. The Raw checkpoint is the pre-anneal weights for continued pretraining; Zero is the 128K base after reinforcement learning, with no supervised fine-tuning stage. These are research artifacts, not products, and have not been safety-evaluated.

PiVo

New

A complete English text-to-speech stack in 1,401,562 learned parameters: a 584,965-parameter acoustic model and an 816,597-parameter vocoder derived from FARGAN, producing 16 kHz audio in two synthetic voices. It runs on a plain CPU without PyTorch, and PiVo Studio runs it from one Rust binary. The words come through clearly; some sentences still sound metallic.

ModelParametersFormatsDownloadsReleased
PiVo-1.4M1.4MONNXNewNovember 2026

Qwen3.8 Distill

7.4M downloads

Full-parameter distillations of the Qwen3.8 frontier models into the Qwen3.5 dense and Qwen3.6 mixture-of-experts architectures. Every answer opens with a reasoning block learnt from teacher traces weighted towards mathematics and code. Native function calling and a 262,144-token context.

ModelParametersFormatsDownloadsReleased
Qwen3.8-35B-A3B-Distill35B total, 3B activeSafetensors, GGUF1,997,270September 2026
Qwen3.8-9B-Distill9BSafetensors, GGUF1,733,737August 2026
Qwen3.8-4B-Distill4BSafetensors, GGUF2,037,972August 2026
Qwen3.8-2B-Distill2BSafetensors, GGUF1,634,999August 2026

Qwen3.8-27B-Ridge

1.1M downloads

A quantization written for Qwen3.8's hybrid layers: the Gated-DeltaNet state path at Q8_0, the mixers at Q4_K, the bits saved in mid-stack feed-forward layers. Nothing was stripped to make the file fit; the native MTP draft head stays in the GGUF. Wiki perplexity 7.82 against 7.15 in BF16.

ModelParametersFormatsDownloadsReleased
Qwen3.8-27B-Ridge-GGUF27B, 11.7 GiBGGUF1,133,296August 2026

Qwythos

6.8M downloads

Full-parameter reasoning models on the Qwen3.5 9B and 27B bases, trained on the same curriculum of verified chain-of-thought traces. The native multi-token-prediction head, the vision tower and a 1,048,576-token context stay intact.

ModelParametersFormatsDownloadsReleased
Qwythos-27B-v127BSafetensors, GGUF885,011July 2026
Qwythos-9B-v29BSafetensors, GGUF1,949,297July 2026
Qwythos-9B-Claude-Mythos-5-1M9BSafetensors, GGUF3,973,274June 2026

Brewed with Brewery

Demo models

Each made start to finish in one guided session with Brewery on a single rented GPU, model card included.

ModelTypeAboutDownloadsReleased
Homebrew-Qwen3.5-2B-Grandmas-Kitchen2BWrites home-cooking recipes in a warm grandmother's voice. Made in one guided session.38October 2026
Homebrew-Qwen-Image-2.1-Y2KImage LoRAA Qwen-Image 2.1 LoRA for the look of early-2000s digital snapshots.500October 2026

Earlier releases

March to June 2026

ModelParametersFormatsDownloadsReleased
Qwable-9B-Claude-Fable-59BSafetensors, GGUF143,060June 2026
Qwen3.5-9B-Claude-Code9BSafetensors, GGUF28,046March 2026
Qwen3.5-9B-Claude-Opus-4.6-Distill9BSafetensors, GGUF12,109March 2026
openNemo-9B9BSafetensors8,809March 2026
openNemo-9B-abliterated9BSafetensors, GGUF15,729March 2026
openNemo-9B-Claude-Opus-4.6-distill9BSafetensors7,685March 2026
openNemo-Cascade-2-30B-A3B30B total, 3B activeSafetensors7,801March 2026

Datasets

10 on Hugging Face

Training traces and task lists used for our distillations, published on Hugging Face.