ModelCloud @ModelCloudAi

Github: https://t.co/QfUNDk4J5Q 🤗 https://t.co/YgYXzLjEmv Our mission is to make ai accessible to any human and robot. {In Stealth-Mode} Earth - Epoch 2.0 Joined June 2024

Tweets

237
Followers

337
Following

168
Likes

210

Qubitium @qubitium

a month ago

🚀 For quant nerds & open-model enjoyers 🧠⚡ GPT-QModel 7.1 main branch just added quantization support for PoolsideAI’s brand-new Laguna models 🌊🏝️ Huge congrats to the Poolside AI team for these open agent-ready models 🤖🔥 @poolsideai @eisokant @jasoncwarner

2 1 8 387 1

View Details

ModelCloud @ModelCloudAi

a month ago

👏GPT-QModel v7.0 major release now on pypi. Now 150x smaller footprint due to just-in-time kernel compilation. 👀

Qubitium @qubitium

a month ago

🚀 GPT-QModel v7.0.0 is here! 🔥 Huawei Ascend NPU support ⚡ JIT compilation for CUDA/ROCm kernels on first use ✅ Easier Pip/UV installs with no --no-build-isolation needed 🧠 Models Added: ⚡ GLM 5 / 5.1 👁️ GLM OCR 🎙️ GLM ASR 💎 Gemma 3n 🦅 Falcon Mamba 🖼️ InternVL Chat

6 2 21 2K 2

0 0 1 99 0

View Details

Qubitium @qubitium

a month ago

Proud to have worked with Nvidia researchers in early 2025 and assisted in implementation and validation of EoRA pipeline in GPT-QModel. It's the real deal! Will be integrating the fused kernel asap.

Min-Hung (Steve) Chen @CMHungSteven

a month ago

🚀EoRA accepted to two @iclr_conf 2026 workshops: @ICBINBWorkshop & #TTU! Fine-tuning-free compensation for compressed LLMs using Eigenspace Low-Rank Approximation. Recover up to +11.45% accuracy on 3-bit quantized models — in just minutes with only hundreds of calibration

3 11 31 5K 9

1 3 2 422 0

View Details

Qubitium @qubitium

2 months ago

🥳 Evalution v0.0.7 released with Tinygrad llm (gguf) inference support. You can now pair Evalution with @__tinygrad__ and use the 153+ and growing llm benchmarks (such as GSM8K Platinum) to test for model quality and inference regressions.

2 2 22 2K 6

View Details

ModelCloud @ModelCloudAi

2 months ago

Regular Expression go zoom zoom!

Qubitium @qubitium

2 months ago

PyPcre v0.3.0 is out 🚀 pip install pypcre 🔥 Python bindings for PCRE2 with full feature coverage ⚡ JIT-fast and faster than re and regex on real workloads 🧵 Thread-safe with nogil support

1 0 2 186 0

0 0 0 49 0

View Details

ModelCloud @ModelCloudAi

2 months ago

🇫🇷 Evalution révolution!

Qubitium @qubitium

2 months ago

🥳 Evalution now on PyPi 👐 🧐Evolve your LLMs with better evals. 🇫🇷 Viva la Evalution! Run with Transformers, vLLM, SGLang, TensorRTLLM, GPTQModel & OpenVINO! ✨ 126 benchmark families. 213 CUDA regression specs. Fast, clean, extensible evals. 🚀

3 6 12 2K 5

0 0 0 47 0

View Details

ModelCloud @ModelCloudAi

2 months ago

Everyone, and we mean everyone benefited greatly from this one man's work. 😍

Qubitium @qubitium

2 months ago

Do I have this right? @tri_dao @togethercompute

0 0 2 338 0

0 0 0 48 0

View Details

Qubitium @qubitium

2 months ago

🦖3 bugs/issues found (one critical) in Transformers Continuous Batching (cb) and FlashAttention2 integration: Wrong cuda graph replay (bad key) causing silent quality collapse. Following PR fixed 2 bugs/issues. The 3rd bug (crash) needs to be fixed separately. @remi_or_ @art_zucker

2 1 4 352 0

View Details

ModelCloud @ModelCloudAi

2 months ago

Lean and fast. The way software is meant to be build. GPT-QModel is now 150x smaller. It will only jit compile the kernels you actually use.

Qubitium @qubitium

2 months ago

GPT-QModel v6.1.0-dev main branch just landed jit compilation for all bundled kernel. 😤 No more 150MB+ plus 66+ variations of python/cuda/torch wheels! 🥳 Best of all, for the first time, you do not need to pass `--nobuild-isolation` to pypi/uv!

0 0 1 396 0

0 0 0 55 0

View Details

ModelCloud @ModelCloudAi

2 months ago

🥳 GPT-QModel v6.0 release is HUGE!

Qubitium @qubitium

2 months ago

🚀 GPTQModel v6.0.3 is out! ✨ ParoQuant added 🎯 FOEM added 🧠 Gemma4, MiniCPM-O, MiniCPM-V 🧠 GLM4-MoE-Lite 🌿 PrismML/Bonsai inference support Easy install: pip install -U gptqmodel --no-build-isolation 🧵👇

1 0 4 583 0

0 0 0 32 0

View Details

ModelCloud @ModelCloudAi

2 months ago

👀Bonsai 1bit model inference support just dropped on GPT-QModel!

Qubitium @qubitium

2 months ago

🥳@PrismML quantized 1bit Bonsai inference support has been added to GPT-QModel v6.0. In addition, I have evaled the 1bit quant and it is excellent for the bit-size. Score on part with the latest 8B models! Wow. What PrimML did was not an April 1st Joke and is the real deal.

1 2 5 635 0

0 0 0 79 0

View Details

ModelCloud @ModelCloudAi

2 months ago

🥳 Official FOEM quantization support added to GPT-QModel main branch.

Qubitium @qubitium

2 months ago

😍 GPT-QModel just landed official support for FOEM: First-Order Error Matters, Accurate Compensation for Quantized LLM. 👇

1 0 2 245 0

0 0 0 47 0

View Details

Qubitium @qubitium

2 months ago

ParoQuant implementation in GPT-QModel is now in beta2 phase and getting close to rc quality. There are still variances vs z-lab official but I am closing the gap fast. GPT-QModel will also offer many levels of customizablee ParoQuant control far and beyond what official provides. It's going to be good.

0 1 2 154 0

View Details

Qubitium @qubitium

2 months ago

openbmb/MiniCPM-o-4_5 support added to GPT-QModel main branch.

0 1 1 168 0

View Details

ModelCloud @ModelCloudAi

2 months ago

GPTQ-QModel now has world class support for Z-Lab's ParoQuant! Benchmarks to come...

Qubitium @qubitium

2 months ago

ParoQuant from Z-lab (UCSD) on GPT-QModel is now getting very good results. Quant speed is now much faster than official while maintaining same accuracy. Thanks to Z-Lab staff feedback, we now have clamping support that can maintain accuracy even with high sampling/epochs.

0 0 1 282 0

0 0 1 61 0

View Details

Qubitium @qubitium

2 months ago

🥳LogBar v0.4.1 patch fix released. Can your py logger do this? Tech demo (unit test) 👇 * Progress bar * Stacked Progress bar * Split regions * Table formatting * Zero dependency and much more...

1 1 1 114 0

View Details

Qubitium @qubitium

5 months ago

👀 Our testing showed that brute forcing activations to all expert modules have negative effect (friendly fire) for all experts. GPT-QModel's naive rtn+mad smoothing in comparison shows no such regression for MoE quants while using much less compute during quantization. 👀

Qubitium @qubitium

5 months ago

🎉 New FailSafe config and FailSafeStrategy, auto enabled by default, to address un-even routing of MoE experts resulting quantization issue of some MoE modules. 👇 Smooth operations are also introduced to FailSafeStrategy to reduce outliers impact... github.com/modelcloud/gpt…

1 0 2 621 0

0 1 2 139 0

View Details

ModelCloud @ModelCloudAi

5 months ago

🔥MoE quantization is no longer a calibration nightmare with latest update from GPT-QModel. 👀

Qubitium @qubitium

5 months ago

1 0 2 621 0

0 0 0 39 0

View Details

ModelCloud @ModelCloudAi

7 months ago

🎆4x faster cpu inference on Intel Xeon contributed by Intel AI just got merged to GPT-QModel. 👇

Qubitium @qubitium

7 months ago

🚀 Do you want 4x faster cpu inference of gptq quantized models on Intel Xeon? 🤯 Intel just dropped a torch compile compatible patch for GPT-QModel's `torch_fused` cpu kernel. 🤗 @IntelAI github.com/ModelCloud/GPT…

0 0 2 168 0

0 0 0 60 0

View Details

ModelCloud @ModelCloudAi

7 months ago

GPT-QModel main branch has fixed AWQ quantization of various MoE models. 👇

Qubitium @qubitium

7 months ago

🥳GPT-QModel dev update: AWQ MoE quantization fixed. Required the creation of a new 🧲:? (? for fishing hook) capture only notation for my high level model definition language (mdl) implemented as a simple tree. Since v5.2 GPTQ and AWQ, and all other supported quant algos,