This repository provides optimized serving configurations for the Qwen3.8-27B model running on a single 24 GB consumer GPU (RTX 3090). It achieves high throughput of approximately 1,000 tok/s at 64 concurrent users in batch mode and up to ~133 tok/s for single-user scenarios using speculative decoding techniques like MTP or DFlash2. The project includes custom patches, requantization scripts (such as int8 tensor-core GEMMs), and Docker support to enable extended context windows of 150k to 262k tokens on a single consumer card.
- Offers two distinct serving profiles: `batch` for high throughput/API backends and `single-user` for low-latency chat experiences.
- Implements advanced speculative decoding modes including MTP (Multi-Token Prediction) and DFlash2 block drafting.
- Supports extreme context lengths up to 262k tokens through KVarN, which utilizes a lossy 4/2-bit KV cache.
- Includes specialized optimizations like int8 activations, quantized lm_head, and split-KV verify attention to maximize VRAM efficiency.
- Offers two distinct serving profiles: `batch` for high throughput/API backends and `single-user` for low-latency chat experiences.
- Implements advanced speculative decoding modes including MTP (Multi-Token Prediction) and DFlash2 block drafting.
- Supports extreme context lengths up to 262k tokens through KVarN, which utilizes a lossy 4/2-bit KV cache.
- Includes specialized optimizations like int8 activations, quantized lm_head, and split-KV verify attention to maximize VRAM efficiency.
Unsloth provides documentation and support for running the GLM-5.3-Flash (ox-alpha) multimodal model locally using Unsloth Desktop or llama.cpp. Developed by Z.ai, this 320B parameter model features a hybrid sparse and linear attention architecture designed to improve scaling through Manifold-Constrained Hyper Connections. Users can utilize various quantization levels'' from 1-bit for low RAM requirements (approx. 93GB) up to higher bitrates for improved accuracy'' to run the model on hardware ranging from Mac systems to NVIDIA DGX Spark setups.
- The model features three thinking modes: Low, High, and Max reasoning effort.
- It is designed to rival Claude Opus 4.8 in coding and agentic benchmarks.
- Unsloth's dynamic 1-bit quantization retains 71% of top-1% accuracy while being 85% smaller than the BF16 version.
- The model can be run via a local API using `unsloth run` with llama-server runtime flags.
- The model features three thinking modes: Low, High, and Max reasoning effort.
- It is designed to rival Claude Opus 4.8 in coding and agentic benchmarks.
- Unsloth's dynamic 1-bit quantization retains 71% of top-1% accuracy while being 85% smaller than the BF16 version.
- The model can be run via a local API using `unsloth run` with llama-server runtime flags.
The Unsloth documentation provides a guide for running Qwen3.8 models locally, covering the 27B vision-and-reasoning model with 256K context that runs on about 17GB VRAM and the massive 2.4T-A95B model with rival GPT-5 class performance. It details recommended thinking vs instruct sampling settings, the Preserve Thinking feature for continued conversations, and hardware requirements across quantizations from 4-bit to 1-bit. The guide explains running models via Unsloth Desktop with automatic offloading and multi-GPU detection, and via llama.cpp with custom builds for IQ1_XXXS quants, plus NVFP4 options for Blackwell GPUs.
- Qwen3.8-27B GGUFs include Developer Role Support for agentic tools like Codex and MTP enabled for fast inference
- Full-precision Qwen3.8-2.4T-A95B needs 4.9TB storage, reduced to 397GB with Unsloth Dynamic 1-bit GGUF ~91% smaller
- NVFP4 quants run ~1.5x faster than BF16 on Blackwell GPUs with 92-97% top-1 accuracy retention
- New 1-bit types TQ1_0/Q1_0 extend IQ1_S to 1.1875 bpw for large models without QAT/QAD
- Qwen3.8-27B GGUFs include Developer Role Support for agentic tools like Codex and MTP enabled for fast inference
- Full-precision Qwen3.8-2.4T-A95B needs 4.9TB storage, reduced to 397GB with Unsloth Dynamic 1-bit GGUF ~91% smaller
- NVFP4 quants run ~1.5x faster than BF16 on Blackwell GPUs with 92-97% top-1 accuracy retention
- New 1-bit types TQ1_0/Q1_0 extend IQ1_S to 1.1875 bpw for large models without QAT/QAD
Qwen Team writes that Qwen3.8-27B is presented as the most capable generation in the Qwen open-model family so far, built on the Qwen3.5 foundation with substantial gains across coding, professional work, research and long-horizon agentic tasks. The model is a 27B-parameter dense causal language model with a vision encoder, native 262,144-token context extensible to 1,000,000 tokens, flexible thinking control with reasoning_effort and preserve_thinking, and Multi-Token Prediction for faster inference. The Hugging Face page hosts Unsloth's GGUF quantizations and provides install and run instructions for llama.cpp, Ollama, Unsloth Studio, LM Studio and other local apps.
This guide outlines the most effective approach for running large language models locally on hardware with 24GB of VRAM. It advises moving away from squeezing extremely large parameter models toward using high-performance 20B to 35B class models that allow room for context and fast processing speeds. The article explains how memory is allocated across model weights, KV cache, and runtime overhead while recommending specific top performers:
* Qwen3.6-27B for agentic coding
* Qwen3.6-35B-A3B MoE for speed in general conversation
* Gemma 4 26B for multimodal and multilingual support
* Mistral Small 3.2 24B as a low-latency assistant
* gpt-oss-20b for structured reasoning tasks
* DeepSeek-R1-Distill-Qwen-32B for deep logical reasoning through chain of thought
* Qwen3.6-27B for agentic coding
* Qwen3.6-35B-A3B MoE for speed in general conversation
* Gemma 4 26B for multimodal and multilingual support
* Mistral Small 3.2 24B as a low-latency assistant
* gpt-oss-20b for structured reasoning tasks
* DeepSeek-R1-Distill-Qwen-32B for deep logical reasoning through chain of thought
This edition of The Weekly Kaitchup reviews several NVFP4 quantization versions of the Qwen3.6 27B model, comparing NVIDIA's mixed-precision approach with community alternatives like Unsloth and PrismaQuant. It also details DSpark, DeepSeek's new speculative decoding method that uses a parallel draft backbone and a confidence head to significantly accelerate large language model generation speeds.
* Comparison of Qwen3.6 27B NVFP4 quantization variants
* Guidance on selecting models based on accuracy versus memory footprint
* Technical overview of DSpark architecture and suffix decay mitigation
* Performance improvements and vLLM support for DSpark
* Comparison of Qwen3.6 27B NVFP4 quantization variants
* Guidance on selecting models based on accuracy versus memory footprint
* Technical overview of DSpark architecture and suffix decay mitigation
* Performance improvements and vLLM support for DSpark
This guide provides instructions for running Alibaba's Qwen3.6 multimodal hybrid-thinking models locally using Unsloth tools. It covers the 27B and 35B-A3B variants, which support a 256K context window across 201 languages and excel in agentic coding, vision, and chat tasks. The article details hardware requirements for various quantization levels and explains how to leverage Multi Token Prediction (MTP) for significantly faster inference.
Key topics:
- Hardware memory requirements for quantized models
- Faster generation via Multi Token Prediction (MTP)
- Integration with Unsloth Studio, llama.cpp, and MLX
- Preserved thinking mode configurations
Key topics:
- Hardware memory requirements for quantized models
- Faster generation via Multi Token Prediction (MTP)
- Integration with Unsloth Studio, llama.cpp, and MLX
- Preserved thinking mode configurations
Turbovec is an open-source vector index library written in Rust that features Python bindings. It utilizes Google's TurboQuant algorithm to provide highly efficient data quantization without the need for traditional codebook training steps like k-means. The library offers significant memory savings, reducing a 31 GB corpus of 10 million vectors down to just 4 GB, and demonstrates superior search speeds on ARM hardware compared to FAISS.
This article explores the feasibility of running Large Language Models (LLMs) locally using only a CPU, challenging the assumption that expensive GPUs are strictly necessary. By testing eight different models on an older Intel i5 laptop with 12GB of RAM via Ollama, the author identifies which models offer practical usability for everyday tasks.
Key points include:
- Using tokens per second as a more critical metric for usability than model size or RAM usage alone.
- Why 1B to 2B parameter models provide the best balance of responsiveness and reasoning on low-end hardware.
- The effectiveness of GGUF quantization (specifically Q4_K_M) in reducing resource demands.
- A comparison of various model tiers, from ultra-fast tiny models like Qwen 0.6B to slower, high-capability models like Ministral 3 8B.
Key points include:
- Using tokens per second as a more critical metric for usability than model size or RAM usage alone.
- Why 1B to 2B parameter models provide the best balance of responsiveness and reasoning on low-end hardware.
- The effectiveness of GGUF quantization (specifically Q4_K_M) in reducing resource demands.
- A comparison of various model tiers, from ultra-fast tiny models like Qwen 0.6B to slower, high-capability models like Ministral 3 8B.
>"One scale parameter determines accuracy in rotation-based vector quantization."
The article demonstrates how the earlier EDEN quantization method outperforms its "successor" TurboQuant by utilizing an analytically optimized scale factor for superior accuracy and bias correction.
* EDEN outperforms newer TurboQuant algorithms.
* Optimal scaling is a key differentiator.
* EDEN-biased minimizes reconstruction error (MSE).
* EDEN-unbiased ensures highly accurate estimation.
* Superior efficiency at low bit-widths.
* Ideal for LLM and KV cache optimization.
The article demonstrates how the earlier EDEN quantization method outperforms its "successor" TurboQuant by utilizing an analytically optimized scale factor for superior accuracy and bias correction.
* EDEN outperforms newer TurboQuant algorithms.
* Optimal scaling is a key differentiator.
* EDEN-biased minimizes reconstruction error (MSE).
* EDEN-unbiased ensures highly accurate estimation.
* Superior efficiency at low bit-widths.
* Ideal for LLM and KV cache optimization.
