# Qwen3.8-27B-NVFP4-RTX5090 by Gittensor Model Hub: Open Model Source: https://savrn.com/models/qwen3-8-27b-nvfp4-rtx5090 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve Qwen3.8-27B-NVFP4-RTX5090 (14.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 29.1 GB | 34.9 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 8-bit | 14.6 GB | 17.5 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 7.3 GB | 8.7 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 25, 2026. [Qwen3.8-27B-NVFP4-RTX5090 on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/qwen3-8-27b-nvfp4-rtx5090/gpus) ## SAVRN's Notes on Qwen3.8-27B-NVFP4-RTX5090 The name says 27B and the page reads 14.6B — confirm which checkpoint you are loading before you benchmark. What is not in doubt: 17.9 GB of NVFP4 weights in two files, 8.7 GB at 4-bit on our table, and one MI300X at $1.85 an hour as the cheapest rental. The build's actual target is a single 32 GB GeForce RTX 5090, with configs shipped for SparkInfer, SGLang and vLLM and the full 262,144-token context claimed to fit. This is an Apache-2.0 repack of Qwen/Qwen3.8-27B aimed at consumer hardware; a DSpark drafter sibling sits in the same hub if you want the publisher's speculative-decoding pairing — the claimed 264.8 tokens a second, up to 420 on code, are publisher numbers, not ours. For an operator the tradeoff is simple: pay data-center rental rates for a 16-bit run, or own the card and run NVFP4 with a vLLM that supports the format. SAVRN Research ## Model Card By Gittensor Model Hub, published under apache-2.0, revision 5b7a687fc821. ### RTX5090 Blackwell GPU Optimized Qwen-3.8-27B-NVFP4 Runs on SparkInfer, SGLang and vLLM unmodified — configs for all three are below. Serving many users at once? See concurrency. ### 420 tok/s writing code. One consumer GPU. SparkInfer × this NVFP4 build × the DSpark v2 drafter — an engine, a checkpoint, and a speculative drafter optimized against each other, compounding to 4.3×. The drafter never changes what the model says: the target verifies every drafted token. GeForce RTX 5090–specific NVFP4 checkpoint of [Qwen/Qwen3.8-27B](https://savrn.com/models/qwen3-8-27b), quantized with [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer). Serves the full native 262,144-token context on 32 GB. With the [DSpark v2 drafter](https://huggingface.co/gittensor-model-hub/Qwen3.8-27B-DSpark-NVFP4): 264.8 tok/s overall — up to 420 on code — on [SparkInfer](https://github.com/gittensor-ai-lab/sparkinfer) (its bench harness; the HTTP server is autoregressive-only today) and 161.7 tok/s on SGLang's OpenAI server. Without speculation: 92.9 tok/s SparkInfer · 85.8 SGLang. [Read the full model card (3,479 words)](https://savrn.com/models/qwen3-8-27b-nvfp4-rtx5090/card) ## Configuration Architecture Qwen3_5ForConditionalGeneration Context length (tokens) 262,144 Layers 64 Hidden size 5,120 Feed-forward size 17,408 Attention heads 24 Key/value heads 4 Head dimension 256 Vocabulary size 248,320 Model type qwen3_5 Quantization modelopt ## Identity and Version Repository gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 Publisher Gittensor Model Hub Task Image and text to text Modality Image and text Library transformers Parameters 14.6B parameters Languages rtx-5090 Revision 5b7a687fc8211a5d631c8ca6a593dd37eb26ce33 First published 2026-08-14 Last updated 2026-09-10 ## Files and Weights 32 files, 17.9 GB in total. The weights are 2 files totalling 17.9 GB in safetensors. Weights2 files · 17.9 GB Configuration7 files · 260.9 KB Tokenizer4 files · 22.9 MB Documentation3 files · 105.2 KB Other15 files · 3.6 MB Repository1 file · 2.2 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model-00001-of-00002.safetensors | Weights | 10.0 GB | cdd37b0e61ec | | model-00002-of-00002.safetensors | Weights | 7.9 GB | 713b84b8287e | | config.json | Configuration | 13.2 KB | — | | generation_config.json | Configuration | 213 B | — | | hf_quant_config.json | Configuration | 9.1 KB | — | | model.safetensors.index.json | Configuration | 236.5 KB | — | | preprocessor_config.json | Configuration | 390 B | — | | processor_config.json | Configuration | 1.2 KB | — | | video_preprocessor_config.json | Configuration | 385 B | — | | LICENSE | Documentation | 11.5 KB | — | | README.md | Documentation | 28.6 KB | — | | README.qwen-upstream.md | Documentation | 65.0 KB | — | | assets/rtx5090-accuracy.png | Other | 86.0 KB | — | | assets/rtx5090-context-vs.png | Other | 230.8 KB | c10600f7439d | | assets/rtx5090-context.png | Other | 81.9 KB | — | | assets/rtx5090-decode.png | Other | 88.7 KB | 2ad1c493144c | | assets/rtx5090-engines.png | Other | 265.7 KB | 6edd0152d89d | | assets/rtx5090-hero-engines.png | Other | 309.0 KB | c7599484f01e | | assets/rtx5090-hero-final.png | Other | 1.5 MB | 8cc0baee6979 | | assets/rtx5090-hero.png | Other | 123.7 KB | 65fdb7d51c26 | | assets/rtx5090-overview.png | Other | 152.1 KB | 798041c149f0 | | assets/rtx5090-spec.png | Other | 122.5 KB | 96d4caee2038 | | assets/rtx5090-speculation.png | Other | 171.6 KB | ac404b7bc6ad | | assets/rtx5090-ttft.png | Other | 71.9 KB | — | | assets/rtx5090-vs-competitors.png | Other | 312.2 KB | bb3915db1add | | chat_template.jinja | Other | 14.2 KB | — | | crc32.txt | Other | 86 B | — | | .gitattributes | Repository | 2.2 KB | — | | merges.txt | Tokenizer | 3.4 MB | — | | tokenizer.json | Tokenizer | 12.8 MB | 0997f410c57a | | tokenizer_config.json | Tokenizer | 1.1 KB | — | | vocab.json | Tokenizer | 6.7 MB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 17.9 GB [Download from Gittensor Model Hub](https://huggingface.co/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090) Released by Gittensor Model Hub through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Derived from [Qwen/Qwen3.8-27B](https://savrn.com/models/qwen3-8-27b) - Quantized from [Qwen/Qwen3.8-27B](https://savrn.com/models/qwen3-8-27b) ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 17.9 GB | | 16-bit | 29.1 GB | | 8-bit | 14.6 GB | | 4-bit | 7.3 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Questions About Qwen3.8-27B-NVFP4-RTX5090 ### How much GPU memory does Qwen3.8-27B-NVFP4-RTX5090 need? About 34.9 GB at 16-bit and 8.7 GB at 4-bit: the weights (14.6B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run Qwen3.8-27B-NVFP4-RTX5090 on? At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### Can I use Qwen3.8-27B-NVFP4-RTX5090 commercially? Yes. Qwen3.8-27B-NVFP4-RTX5090 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. ### What is Qwen3.8-27B-NVFP4-RTX5090's context length? 262,144 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Image and text to text ### [Swift-1.5-Qwen3.8-27b-Quark-RTN-MXFP4](https://savrn.com/models/swift-1-5-qwen3-8-27b-quark-rtn-mxfp4) [Ethan Todd](https://savrn.com/model-publishers/ethantodd4l) quantized to OCP MXFP4 4-bit weights in the same Quark checkpoint container as The weights are plain RTN, not AWQ: see the note below. The 15 MTP tensors are kept in BF16; the radiance runtime loads them with RADIANCEQUARKBF16MTP=1. Why RTN instead of AWQ. The first build of this checkpoint used AWQ with the same smoothing recipe as AMD's release. The AWQ fold itself was mathematically consistent, but on this model a few layers converged to weights to 32-element MXFP4 blocks then destroyed those layers on real, outlier-carrying inputs (layer 7 output relative error ~324), and repairing the worst layers individually was not enough; the remaining smoothed layers still accumulated too much… Open weights other 15.6B parameters 262,144 tokens transformers [View model](https://savrn.com/models/swift-1-5-qwen3-8-27b-quark-rtn-mxfp4) Model · Image and text to text ### [span-02](https://savrn.com/models/span-02) [He Li](https://savrn.com/model-publishers/lihe07) span-02 is a decision model: it answers typed questions about a state (JSON or text, plus up to 4 images) with a full probability distribution over the options, in the Decision Index /v1/systemone wire format. It is a fine-tune of Gemma 4 26B-A4B (mixture of experts, about 4B active parameters per token), quantized (NVFP4 experts, FP8 dense layers) and served by a custom Rust/CUDA engine for Blackwell GPUs (sm120: RTX 5090, RTX PRO 6000) that is included in this repo (engine/). Before answering it may think. If the top probability of the no-think answer is below 0.8, the model writes a greedy thought. Its length limit comes from a deterministic thought budget: a fixed cost model… Open weights apache-2.0 15.8B parameters 262,144 tokens [View model](https://savrn.com/models/span-02) Model · Image and text to text ### [SDRPN-Gemma-4-12B](https://savrn.com/models/sdrpn-gemma-4-12b) [Yuheng Shi](https://savrn.com/model-publishers/yuhengsss) Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN) SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 27, T = 3) trained on a frozen, encoder-free google/gemma-4-12B-it. This is the initialisation of the Vision-RL² RL run The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation). The Gemma stage-1 driver trains the delta and its assemble stage turns it into the full directory; both forms are published so you can start from either. These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable… Open weights apache-2.0 12.6B parameters 262,144 tokens [View model](https://savrn.com/models/sdrpn-gemma-4-12b) Model · Image and text to text ### [GLM-5.3-Flash-NestQuant-1.5-4bit](https://savrn.com/models/glm-5-3-flash-nestquant-1-5-4bit) [Jarrel Seah](https://savrn.com/model-publishers/jarrelscy) Run one of these commands on a separate scorer workstation with SSH access to an idle native ARM64 DGX Spark. The launcher securely prompts for the serving API key, downloads and prepares the pinned model and ARM64 Docker image on the Spark, launches the exact accepted 2,630-slot configuration, opens an authenticated SSH tunnel, installs the pinned official scorer, and runs all 80 tasks twice. Reasoning on — reproduces the canonical 89.2151 run configuration: Reasoning off — reproduces the 86.6870 run configuration: Both options were live-tested against the published Spark service: SSH and host-key validation, authenticated tunnel, endpoint/model check, pinned scorer checkout, 73/73… Open weights mit 16.9B parameters 1,048,576 tokens [View model](https://savrn.com/models/glm-5-3-flash-nestquant-1-5-4bit) Model · Image and text to text ### [Jev-Omni-MLX-4bit](https://savrn.com/models/jev-omni-mlx-4bit) [Leo Zhang](https://savrn.com/model-publishers/ruiruiz30) An Apple Silicon MLX conversion of akhilaaa3/Jev-Omni for local inference on a Mac mini with 16GB unified memory. This repository keeps the upstream Jev-Omni unified multimodal checkpoint and its trained 256-way decision head, then converts the language-model weights to 4-bit affine quantization with group size 64. The vision weights remain BF16 and the decision head remains FP32. It is an independent conversion. It is not an official TypeSafe Jev release and does not claim to reproduce TypeSafe's proprietary system. It is also not a new fine-tune. - Mac mini, Apple M4, 10 CPU cores, 16GB unified memory - macOS 26.5.1 - Python 3.13.12 - MLX 0.32.2 - MLX-VLM 0.7.1 - single request, batch… Open weights apache-2.0 12B parameters 262,144 tokens mlx [View model](https://savrn.com/models/jev-omni-mlx-4bit) Model · Image and text to text ### [Clemma-12B](https://savrn.com/models/clemma-12b) [Andrew DeLisa](https://savrn.com/model-publishers/ayan4m1) This gemma4unified model was trained 2x faster with Unsloth and Huggingface's TRL library. Open weights apache-2.0 12B parameters 262,144 tokens transformers [View model](https://savrn.com/models/clemma-12b) ## Gittensor Model Hub [All models and datasets](https://savrn.com/model-publishers/gittensor-model-hub) ## Versions - [5b7a687fc821](https://savrn.com/models/qwen3-8-27b-nvfp4-rtx5090/versions/5b7a687fc821) · current 2026-09-18 ## Explore More - [All image and text to text models](https://savrn.com/models/tasks/image-and-text-to-text) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-19. - [Hugging Face record](https://huggingface.co/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090) - [How the hub is built](https://savrn.com/model-hub/methodology)