Qwen3.5 122B-A10B

Amazon Web Services announced that Qwen3.5-122B-A10B is among the models now available on Amazon SageMaker JumpStart, alongside LocateAnything-3B and Qwen-AgentWorld-35B-A3B. The announcement page explicitly names Qwen3.5-122B-A10B as part of the JumpStart catalog update, giving developers an AWS-managed path to deploy As an Apache 2.0 licensed 122B-parameter MoE with roughly 10B active parameters and a ~256K token context, Qwen3.5-122B-A10B is designed for text and visual reasoning across documents, charts, and natural language. Its inclusion in SageMaker JumpStart lowers the operational barrier for enterprises wanting to run the mo

Page 20 of the same NVIDIA DGX Spark forum thread captures a May 2026 discussion benchmarking vLLM 0.19.1 against 0.20.1 on Qwen3.5-122B-A10B (hybrid INT4+FP8 + INT8 LM Head + MTP-2) on a single Spark. The author reports a ~7.7–10.5% regression across Q&A, Code, JSON, Math, and LongCode workloads, attributing the drop Community contributors note that the Qwen3.5-122B-A10B gdn attention core (DeltaNet linear attention, 36 of 48 layers) is now split out rather than stitched into one graph, and on a 273 GB/s UMA Spark with low SM count each extra boundary is real overhead; on H100/H200 (where vLLM CI runs) the effect is reportedly invi

Community oMLX benchmark results dated May 10, 2026 show Qwen3.5-122B-A10B running at 4-bit quantization on an M4 Max with 128 GB unified memory and 40 GPU cores under oMLX v0.3.8 on macOS 26.4. At a 1K context, it sustains 526.8 prompt tokens/s and 57.5 generation tokens/s with peak memory of 65.6 GB. Throughput declines with context length, dropping to 263.9 PP tok/s and 25.4 TG tok/s at 128K tokens (87.2 GB peak memory). Batching tests on the same M4 Max setup show generation throughput scaling from 57.5 tok/s at batch 1 to 84.8 tok/s at batch 2 (1.47x) and 112.6 tok/s at batch 4 (1.96x speedup). Peak memory climbs modestly with context, from 65.6 GB at 1K to 76.8 GB at 64K, indicating the 4-bit quantized model fits comfortably within 128 GB even at long contexts. These figures are from a single community submission and may not transfer to full-precision deployments.

BenchLM's September 25, 2026 tracker places Qwen3.5-122B-A10B at an overall capability score of 40.1/100 (ranked 113 of 194 models), with the strongest category being instruction following at rank 11 of 124 (score 91.6, 92nd percentile). The model reports a 262K token context and a measured throughput of 129 tok/s against a field median of 91 tok/s, placing it among faster open-weights options. Multilingual performance ranks it 10th of 12 tracked models in that category. Per BenchLM, the model's published evidence includes 15 displayable benchmark rows, with verified scores in agentic (23.1, rank 84/105), coding (35.7, rank 73/135), reasoning (49.8), multimodal (57.0, rank 34/50), and knowledge (41.5, rank 86/158). No comparable first-party hosted token rate is published, and math benchmarks are not measured for this variant. The tracker flags the context-length figure as lacking a stored direct source link.

Qwen3.5-122B-A10B is Alibaba Cloud's mid-tier multimodal MoE foundation model, released February 23, 2026, with 122B total parameters and 10B activated per token across 256 experts (apxml.com specs page). It pairs a 262K native context window, extendable to 1M tokens, with grouped-query attention (32Q/2KV heads, head dim 256) and SwiGLU feed-forward layers. The page lists benchmark scores including MMLU-Pro 0.867 and GPQA 0.866, and recommends 3x RTX 6000 Blackwell or a single MI325X for full FP16 hosting. According to apxml.com, the model is distributed under Apache 2.0 and reports active parameter count of 122B with 450M auxiliary parameters, supporting multimodal inputs across vision and language. A 128K-context FP16 inference workload requires roughly 245.72 GB VRAM, dominated by expert weights. Its strongest published rankings are in knowledge (rank 86 of 158) and coding (rank 96), with an overall catalog rank of 100 among tracked models.

Roboflow's Playground page documents Qwen3.5-122B-A10B as a high-capacity multimodal Mixture-of-Experts model from Alibaba's Qwen team, with 122B total parameters activating approximately 10B per token via sparse expert routing. The page specifies the model supports both text and visual inputs in a unified framework fo The entry includes independent usage telemetry showing 41 inferences in the trailing 30 days with an average latency of 18.02 seconds, and reports a legacy Vision Evals pass rate of 76.12% across 67 tasks, placing the model 9th of 77 in that evaluation and above 86% of peers on that benchmark. Methodology details for t