llmrun
OpenGVLabGPT-OSSInternVLForConditionalGeneration

InternVL3 5 GPT OSS 20B A4B Preview HF — Hardware Requirements & GPU Compatibility

Vision

InternVL3.5-GPT-OSS-20B-A4B-Preview-HF stands as OpenGVLab's multimodal MoE model, featuring roughly 21.2 billion total parameters and 4 billion active ones. This build leans on OpenAI's gpt-oss-20b for language tasks and InternViT-300M for vision. As a preview within the InternVL3.5 line, it ships in Transformers-native HF format. Documentation notes that lmdeploy lacks support for GPT-OSS for now; instead, vLLM is recommended for deployment. With only 4 billion parameters active, inference remains relatively quick. Once quantized, it fits on a single high-end consumer GPU or a machine boasting at least 16 GB of memory. Released under the Apache 2.0 license, it allows both commercial and research use, though the gpt-oss base model stays separate. Published in August 2025, this entry is the MoE variant of a family that otherwise relies on Qwen3-based dense language models.

21.2B params · MoE4.5B active131K contextApache 2.0660.2K downloads10 likes
S

Runs great on a 24 GB GPU

Runs from12 GB Q3_K_M · 32KBest on24 GB Q4_K_M · ~180 tok/s · 128KMacs from16 GB Q3_K_M · ~48 tok/s

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run InternVL3 5 GPT OSS 20B A4B Preview HF

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended

vLLM
one command · needs vLLM installed
$vllm serve OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF --max-model-len 16384

vLLM repo with its own quantisation

Download: Weights on Hugging Face

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX 3090 (24 GB) (10 GB headroom) · Q4_K_M

Generation speed
~170 tok/s
generation speed
Cost per 1M output tokens
$0.32
per 1M output tokens

How Much VRAM Does InternVL3 5 GPT OSS 20B A4B Preview HF Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn RTX 4090 (reference)
9.4 GB10.5 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsFits: 128K context · ~229 tok/s
10.7 GB11.8 GB3-bit medium — 3-bit medium quantizationFits: 128K context · ~209 tok/s
13.1 GB14.2 GB4-bit medium — 4-bit medium quantization — most popular sweet spotFits: 128K context · ~180 tok/s
15.5 GB16.6 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffFits: 128K context · ~158 tok/s
17.9 GB19 GB6-bit, very good — 6-bit quantization, very good qualityFits: 128K context · ~140 tok/s
21.6 GB22.7 GB8-bit, near-lossless — 8-bit quantization, near-losslessFits: 64K context · ~120 tok/s
42.8 GB43.9 GB16-bit · bfloat — Brain floating point 16 — preferred for trainingDoes not fit: needs 43.3 GB
est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

13.6GBat 16K context

Weights
12.7 GB
KV cache
0.6 GB
Overhead
0.3 GB

Fits RTX 4090 (reference) · 10.4 GB free · ~180 tok/s

≥50 tok/s needs ≥229 GB/s · RTX 5090 ~269, RTX 3090 Ti ~180

Which Hardware Can Run InternVL3 5 GPT OSS 20B A4B Preview HF?

Q4_K_M · 13.1 GB

InternVL3 5 GPT OSS 20B A4B Preview HF (Q4_K_M) requires 13.1 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 18+ GB is recommended. Using the full 131K context window can add up to 4.5 GB, bringing total usage to 17.6 GB. 30 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 139 devices with unified memory can run InternVL3 5 GPT OSS 20B A4B Preview HF, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), MacBook Pro 14" M3 Pro (18 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M
8 GB · RTX 4060Doesn't fit — smallest quant needs 9.9 GB8 GB tier →
12 GB · RTX 3060 12GBQ3_K_M · 11.2 GB · ~91 tok/s · 32K12 GB tier →
16 GB · RTX 5060 Ti 16GBQ4_K_M · 13.6 GB · ~92 tok/s · 64K16 GB tier →
24 GB · RTX 4090referenceQ4_K_M · 13.6 GB · ~180 tok/s · 128K24 GB tier →
Mac 32 GB · Mac Mini M4Q4_K_M · 13.6 GB · ~26 tok/s · 128KMacs →
Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 13.6 GB · ~83 tok/s · 128KMacs →

Alternatives to InternVL3 5 GPT OSS 20B A4B Preview HF

GPT-OSS family · 14 models →

Similar models

6
ModelScoreVRAMtok/sWhy it is similar
Internlm 20BInternLM—45.3 GB—needs > 24 GBsame size
Tri 21B Thinktrillionlabs—45.1 GB—needs > 24 GBsame size
EuroLLM 22B Instruct 2512utter-project—17.5 GB~46same size
ERNIE 4.5 21B A3B ThinkingBaidu—14.3 GB~213same size
Mistral Small Instruct 2409Mistral AI—17.4 GB~46same size
GPT Neox 20BEleutherAI—13.7 GB~48same size

Specifications

Publisher
OpenGVLab
Family
GPT-OSS
Parameters
21.2B
Active parameters
4.5B per token
Architecture
InternVLForConditionalGeneration
Context Length
131,072 tokens
Vocabulary Size
200,028
Release Date
2025-08-29
License
Apache 2.0

Frequently Asked Questions

How much VRAM does InternVL3 5 GPT OSS 20B A4B Preview HF need?

InternVL3 5 GPT OSS 20B A4B Preview HF requires 13.1 GB of VRAM at Q4_K_M, or 42.8 GB at BF16. Full 131K context adds up to 4.5 GB (17.6 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 21.2B × 4.8 bits ÷ 8 = 12.7 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 4.9 GB (at full 131K context)

VRAM usage by quantization

13.1 GB
17.6 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run InternVL3 5 GPT OSS 20B A4B Preview HF?

Yes, at Q8_0 (21.6 GB) or lower. Higher quantizations like BF16 (42.8 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for InternVL3 5 GPT OSS 20B A4B Preview HF?

For InternVL3 5 GPT OSS 20B A4B Preview HF, Q4_K_M (13.1 GB) offers the best balance of quality and VRAM usage. Q5_K_M (15.5 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 9.4 GB.

VRAM requirement by quantization

Q2_K
9.4 GB
Q4_K_M ★
13.1 GB
Q5_K_M
15.5 GB
Q6_K
17.9 GB
Q8_0
21.6 GB
BF16
42.8 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run InternVL3 5 GPT OSS 20B A4B Preview HF on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 145 other Macs can run InternVL3 5 GPT OSS 20B A4B Preview HF, the smallest at Q3_K_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 13.1 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).

Can I run InternVL3 5 GPT OSS 20B A4B Preview HF locally?

Yes — InternVL3 5 GPT OSS 20B A4B Preview HF can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is InternVL3 5 GPT OSS 20B A4B Preview HF?

At Q4_K_M, InternVL3 5 GPT OSS 20B A4B Preview HF can reach ~186 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~180 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 2.8 GB active ÷ (8000 × 0.65) = 0.536 ms, plus 24 layers × 0.055 ms = 1.320 ms, so 1000 ÷ 1.856 ms = ~539 tok/s

Estimated speed at Q4_K_M (13.1 GB)

~539 tok/s
~180 tok/s
~539 tok/s
~456 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of InternVL3 5 GPT OSS 20B A4B Preview HF?

At Q4_K_M, the download is about 12.74 GB. The full-precision BF16 version is 42.47 GB. The smallest option (Q2_K) is 9.02 GB.

Which GPUs can run InternVL3 5 GPT OSS 20B A4B Preview HF?

30 consumer GPUs can run InternVL3 5 GPT OSS 20B A4B Preview HF at Q4_K_M. Top options include AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, AMD Radeon RX 6800. 8 GPUs have plenty of headroom for comfortable inference.

Which devices can run InternVL3 5 GPT OSS 20B A4B Preview HF?

139 devices with unified memory can run InternVL3 5 GPT OSS 20B A4B Preview HF at Q4_K_M (13.1 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.