Model

Qwen3.5-122B-A10B

OPEN

Alibaba · Qwen3.5 family · released Feb 24, 2026

Mid-size open-weight MoE in the Qwen3.5 family, released ~1 week after the 397B-A17B flagship alongside Qwen3.5-27B (dense) and Qwen3.5-35B-A3B (MoE). Matches or beats GPT-OSS-120B and Qwen3-235B-A22B on several benchmarks at a fraction of the total/active parameter count.

ReasoningCodingVisionFunction callingTool useAgentic

2004.1

Elo · rank #77

Parameters

122B

Active params

10B (MoE)

Context

256K tokens

Architecture

Hybrid-attention MoE (Qwen3-Next lineage): 48 layers, 12x(3x(Gated DeltaNet -> MoE) -> 1x(Gated Attention -> MoE)); 256 experts, 8 routed + 1 shared activated per token; hidden dim 3072; natively multimodal (Causal LM with built-in vision encoder)

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision93.3%#14

best: Molmo 72B · 96.3%

Arena EloHuman preference1417#19

best: Claude Fable 5 · 1505

BrowseCompAgents63.8%#28

best: Kimi K3 · 91.2%

C-EvalKnowledge91.9%#7

best: Qwen3.6-Plus · 93.3%

CharXivVision77.2%#24

best: Qwen3.8-Flash-Next · 90.6%

CodeforcesCoding2100#15

best: DeepSeek-V4-Pro (Think Max) · 3206

GPQA DiamondReasoning86.6%#56

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning25.3%#60

best: Claude Opus 5 · 64.7%

IFBenchReasoning76.1%#16

best: MiniMax M3 · 83.0%

IFEvalReasoning93.4%#7

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding78.9%#56

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision86.2%#8

best: Seed 2.1 Pro · 92.6%

MathVistaVision87.4%#4

best: Seed 2.1 Pro · 90.7%

MMBench (English)Vision92.8%#2

best: Qwen3.5-397B-A17B · 93.7%

MMLU-ProKnowledge86.7%#18

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge94.0%#8

best: Qwen3.7-Max · 95.0%

MMMLUKnowledge86.7%#15

best: Gemini 3.1 Pro · 92.6%

MMMU-ProVision76.9%#24

best: Claude Opus 4.7 · 85.5%

MMMUVision83.9%#11

best: Claude Fable 5 · 89.3%

OCRBenchVision92#15

best: InternVL3-78B · 906

OSWorld-VerifiedAgents58.0%#26

best: Claude Fable 5 · 85.0%

SuperGPQAReasoning67.1%#7

best: Qwen3.7-Max · 73.6%

SWE-bench VerifiedCoding72.0%#56

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding49.4%#62

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA83.0 GB1× H200 141GB

LoRA259.9 GB2× H200 141GB

Full fine-tune1958.1 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $63.09 (1× H200 141GB)

Plan a fine-tune →

Qwen3.5 family

Elo progression across releases

Qwen3.5-397B-A17BFeb 20262146.0 Qwen3.5-PlusFeb 20261864.3 Qwen3.5-122B-A10BFeb 20262004.1 Qwen3.5-Omni-FlashMar 20261454.8 Qwen3.5-Omni-PlusMar 20261817.9

Resources

HFQwen/Qwen3.5-122B-A10B GHQwenLM/Qwen3.6 DOCShttps://huggingface.co/Qwen/Qwen3.5-122B-A10B APIAvailable via Alibaba API

API price weights · each benchmark row carries its own source badge (see methodology)