Skip to content
Updated minutes ago
Alibaba

Qwen 3 235B

Alibaba · moe · 235B parameters · 131,072 context

Quality
83.0

Parameters

235B

Context Window

128K tokens

Architecture

MoE

Best GPU

B200 SXM

Cheapest API

$3.00/M

Quality Score

83/100

Intelligence Brief

Qwen 3 235B is a 235B parameter Mixture-of-Experts (128 experts, 8 active) model from Alibaba, featuring Grouped Query Attention (GQA) with 94 layers and 5,120 hidden dimensions. With a 131,072 token context window, it supports tools, structured output, code, math, multilingual, reasoning. On standardized benchmarks, it achieves MMLU 88, HumanEval 62, GSM8K 94. The most cost-effective API deployment is via together at $3.00/M output tokens. For self-hosted inference, B200 SXM delivers optimal throughput at $8522/month.

Architecture Details

TypeMOE
Total Parameters235B
Active Parameters22B
Layers94
Hidden Dimension5,120
Attention Heads64
KV Heads4
Head Dimension128
Vocab Size151,936
Total Experts128
Active Experts8

Memory Requirements

BF16 Weights

470.0 GB

FP8 Weights

235.0 GB

INT4 Weights

117.5 GB

KV-Cache per Token192512 bytes
Activation Estimate3.00 GB

Fits on (multi-GPU with Tensor Parallelism)

Multi-GPU configurations use Tensor Parallelism (TP) to split model layers across GPUs. Requires NVLink or NVSwitch interconnect for optimal performance.

GPU Compatibility Matrix

Qwen 3 235B is compatible with 8% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
B300288GB
B100 SXM192GB
GB200 NVL72 (per GPU)192GB
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
H200 SXM141GB
H2096GB
GH20096GB
Ada Lovelace(11 GPUs)
L40S48GB
L4048GB
RTX 6000 Ada48GB
L2048GB
Ampere(16 GPUs)
A100 80GB SXM80GB
A100 80GB PCIe80GB
A1664GB
RTX A600048GB
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

B200 SXMoptimal

FP8 · 2 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$8522

Cost/M Tokens

$11.58

Use this config →
B100 SXMoptimal

FP8 · 2 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$8541

Cost/M Tokens

$11.61

Use this config →
GB200 NVL72 (per GPU)optimal

FP8 · 2 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$12337

Cost/M Tokens

$16.77

Use this config →

Deployment Options

API

API Deployment

together

$3.00/M

output tokens

Self-Hosted

Single GPU

B200 NVL (pair)

$9965/mo

Min VRAM: 235 GB

Scale

Multi-GPU

B200 SXM x2

280.0 tok/s

TP· $8522/mo

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
together$1.50$3.00
Cheapest
fireworks$1.80$3.50

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
togetherBest Value$1.50$3.00$23
fireworks$1.80$3.50$27

Cost per 1,000 Requests

Short (500 tok)

$1.35

via together

Medium (2K tok)

$5.40

via together

Long (8K tok)

$18.00

via together

Performance Estimates

Throughput by GPU

B200 SXM
280.0 tok/s
B100 SXM
280.0 tok/s
GB200 NVL72 (per GPU)
280.0 tok/s

VRAM Breakdown (B200 SXM, FP8)

Weights
Act
Weights 117.5 GBKV-Cache 1.6 GBActivations 24.0 GBOverhead 5.9 GB

Precision Impact

bf16

235.0 GB

weights/GPU

fp8

117.5 GB

weights/GPU

~280.0 tok/s

int4

58.8 GB

weights/GPU

Quality Benchmarks

Top 10%
91th percentile across all models
MMLU
88.0
Above Average (86th pctile)
HumanEval
62.0
Average (73th pctile)
GSM8K
94.0
Above Average (84th pctile)
MT-Bench
88.0
Bottom 25% (0th pctile)

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmsglangtensorrt-llm

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Qwen 3 235B

Similar Models

Lower quality, CheaperCompare →

DeepSeek V2.5

236B params · moe

Quality: 78

from $0.28/M

More expensiveCompare →

Claude Opus 4

200B params · dense

Quality: 90

from $75.00/M

Larger context, Higher quality, More expensiveCompare →

GPT-4o

200B params · moe

Quality: 85

from $10.00/M

More expensiveCompare →

Frequently Asked Questions

How much VRAM does Qwen 3 235B need for inference?

Qwen 3 235B requires approximately 470.0 GB of VRAM at BF16 precision, 235.0 GB at FP8, or 117.5 GB at INT4 quantization. Additional VRAM is needed for KV-cache (192512 bytes per token) and activations (~3.00 GB).

What is the best GPU for Qwen 3 235B?

The top recommended GPU for Qwen 3 235B is the B200 SXM (x2) using FP8 precision. It achieves approximately 280.0 tokens/sec at an estimated cost of $8522/month ($11.58/M tokens). Score: 100/100.

How much does Qwen 3 235B inference cost?

Qwen 3 235B API inference starts from $1.50/M input tokens and $3.00/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.