Model reference · open weights

Qwen3.8-RTX5090-LMHead4

LLMs gittensor-model-hub Vision + text 1 build Open weights 12k dl/mo

Qwen3.8-RTX5090-LMHead4 is an open-weight language model from gittensor-model-hub. Qwen3.8-27B-NVFP4-RTX5090-LMHead4 (NVFP4) weighs 18.8 GB; the smallest configuration that runs it is RTX 4090 24 GB.

Request Rent a GPU

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

Qwen3.8-RTX5090-LMHead4 is an image-text-to-text model with 15.0B parameters and a 262144-token context length, released by gittensor-model-hub under the Apache 2.0 license.
This variant quantizes the lm_head to NVFP4, reducing checkpoint size to 18.77 GB and increasing decode speed to 88.45 tokens per second on an RTX 5090.
The card notes that the model is superseded by a newer checkpoint that also removes the MTP head.

Summary of the gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090-LMHead4 model card, 2026-10-01

Hugging Face ↗

What it is

Released by	gittensor-model-hub
Released	2026-08-19
Parameters	15.0B
VRAM	18.8 GB for the weights

What it runs on

Memory and cards for Qwen3.8-27B-NVFP4-RTX5090-LMHead4 (NVFP4)

18.8 GBweights, file size

33 MBcache per 1K tokens (8-bit)

308 MBfixed state per request

2.2 GBruntime overhead, at least

262,144 tokenscontext max

Card	Requests at once		Context max	Memory
Card	8K each	32K each	Context max	Memory
RTX 3060 12 GB … RTX 4060 Ti 16 GB 2 smaller cards
RTX 3090 24 GB FP4 without its speed-up here	4	1	63K	23.4 GB
RTX 4090 24 GB FP4 without its speed-up here	4	1	62K	23.4 GB
RTX 5090 32 GB	17	7	all 256K	31.0 GB
L40S 48 GB FP4 without its speed-up here	39	16	all 256K	44.0 GB
A100 80 GB FP4 without its speed-up here	99	41	all 256K	78.2 GB
H100 80 GB FP4 without its speed-up here	89	37	all 256K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	116	48	all 256K	93.8 GB
DGX Spark (GB10) 128 GB unified	140	58	all 256K	107 GB
H200 141 GB FP4 without its speed-up here	193	80	all 256K	138 GB
B200 180 GB	258	107	all 256K	176 GB
2× RTX 4060 Ti 16 GB tensor parallel · FP4 without its speed-up here	13	5	219K	15.4 GB a card
2× RTX 4090 24 GB tensor parallel · FP4 without its speed-up here	40	17	all 256K	23.4 GB a card
2× RTX 3090 24 GB tensor parallel · FP4 without its speed-up here	41	17	all 256K	23.4 GB a card

Memory needed at each load

Requests at once	8K tokens each	32K tokens each
1	21.5 GB	22.3 GB
5	23.8 GB	27.9 GB
8	25.6 GB	32.0 GB
16	30.2 GB	43.1 GB
32	39.4 GB	65.2 GB
64	57.8 GB	109 GB

One card, with vLLM's small-card settings.

Details

Type	Language models
Task	Vision + text
Context	262,144 tokens
Runs with	transformers
Based on	gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090
Popularity	12k downloads / month
Weights	18.8 GB, Qwen3.8-27B-NVFP4-RTX5090-LMHead4 (NVFP4, file size)