Updated minutes ago
ReleasedJuly 21, 2025Verified — · huggingface.co
Alibaba

Qwen3-235B-A22B-Thinking-2507

Qwen · moe · 235B parameters · 262,144 context

Quality
50.0
Deploy this model →Use via API →Calculate ROI →Compare with othersFine-tune this model

Parameters

235B

Context Window

256K tokens

Architecture

MoE

Best GPU

B200 NVL (pair)

Cheapest API

$0.60/M

Intelligence Brief

Qwen3-235B-A22B-Thinking-2507 is a 235B parameter Mixture-of-Experts (128 experts, 8 active) model from Qwen, featuring Grouped Query Attention (GQA) with 94 layers and 4,096 hidden dimensions. With a 262,144 token context window, it supports tools, structured output, code, math, multilingual, reasoning. The most cost-effective API deployment is via openrouter at $0.60/M output tokens. For self-hosted inference, B200 NVL (pair) delivers optimal throughput at $19929/month.

Provider pricing

1 provider · canonical: openrouter
Provider Input $/M Output $/M ▲Notes
openroutercanonical$0.130$0.600cheapest input · cheapest output

Prices update via the nightly pricing cron + admin approvals at /admin/ingest-queue. The leaderboard's Input/Output cells show the canonical rate above; this table shows the full spread.

Recent changes

Loading…

Related models

3 suggestions

Picks: same family first, then same vendor within ±2× params, then top tag-overlap matches. Price shown is the cheapest Output $/M across providers — the row's page shows the canonical anchor.

Architecture Details

TypeMOE
Total Parameters235B
Active Parameters22B
Layers94
Hidden Dimension4,096
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 Estimate0.00 GB

Fits on (single GPU) — most practical first

BEST FITH200 SXM141GBINT4B200 SXM180GBINT4H100 NVL 94GB (per GPU pair)188GBINT4B100 SXM192GBINT4GB200 NVL72 (per GPU)192GBINT4GB300 NVL72 (per GPU)192GBINT4Instinct MI300X192GBINT4Groq LPU230GBINT4Instinct MI325X256GBINT4B300288GBFP8B200 NVL (pair)360GBFP8

Fits on (multi-GPU with Tensor Parallelism)

RECOMMENDEDx2H100 SXM80GB eachINT4

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

GPU Compatibility Matrix

Qwen3-235B-A22B-Thinking-2507 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
✗
35%
67%
B300288GB
✗
18%
59%
B100 SXM192GB
✗
✗
39%
GB200 NVL72 (per GPU)192GB
✗
✗
39%
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
✗
✗
38%
H200 SXM141GB
✗
✗
17%
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 NVL (pair)optimal

BF16 · 2 GPUs · tensorrt-llm

98/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$19929

Cost/M Tokens

$27.08

Use this config →
B200 SXMoptimal

BF16 · 4 GPUs · tensorrt-llm

93/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$17044

Cost/M Tokens

$23.16

Use this config →
B100 SXMoptimal

BF16 · 4 GPUs · tensorrt-llm

93/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$17082

Cost/M Tokens

$23.21

Use this config →

Deployment Options

API

API Deployment

openrouter

$0.60/M

output tokens

Self-Hosted

Single GPU

Requires multi-GPU setup (235 GB VRAM needed)

Scale

Multi-GPU

B200 NVL (pair) x2

280.0 tok/s

TP· $19929/mo

Open Full Calculator →

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
openrouter$0.13$0.60
Cheapest

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
openrouterBest Value$0.13$0.60$4

Cost per 1,000 Requests

Short (500 tok)

$0.18

via openrouter

Medium (2K tok)

$0.74

via openrouter

Long (8K tok)

$2.24

via openrouter

Performance Estimates

Throughput by GPU

B200 NVL (pair)
280.0 tok/s
B200 SXM
280.0 tok/s
B100 SXM
280.0 tok/s

VRAM Breakdown (B200 NVL (pair), BF16)

Weights
Weights 235.0 GBKV-Cache 3.2 GBActivations 0.0 GBOverhead 11.8 GB

Capabilities

Features

✓ Tool Use✗ Vision✓ Code✓ Math✓ Reasoning✓ Multilingual✓ Structured Output

Supported Frameworks

vllmsglang

Supported Precisions

BF16 (default)

Where to Deploy Qwen3-235B-A22B-Thinking-2507

Similar Models

Qwen 3 235B

235B params · moe

Quality: 83

from $0.00/M

Higher quality, CheaperCompare →

DeepSeek Coder V2 236B

236B params · moe

Quality: 50

from $0.28/M

CheaperCompare →

DeepSeek V2.5

236B params · moe

Quality: 78

from $0.28/M

Higher quality, CheaperCompare →

MiniMax-M2.1

229B params · moe

Quality: 50

Similar specsCompare →

MiniMax-M2.5

229B params · moe

Quality: 50

from $1.20/M

More expensiveCompare →

Frequently Asked Questions

How much VRAM does Qwen3-235B-A22B-Thinking-2507 need for inference?

Qwen3-235B-A22B-Thinking-2507 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 (~0.00 GB).

What is the best GPU for Qwen3-235B-A22B-Thinking-2507?

The top recommended GPU for Qwen3-235B-A22B-Thinking-2507 is the B200 NVL (pair) (x2) using BF16 precision. It achieves approximately 280.0 tokens/sec at an estimated cost of $19929/month ($27.08/M tokens). Score: 98/100.

How much does Qwen3-235B-A22B-Thinking-2507 inference cost?

Qwen3-235B-A22B-Thinking-2507 API inference starts from $0.13/M input tokens and $0.60/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.