Qwen3.5 122B-A10B

Community users on the NVIDIA DGX Spark forum report running Qwen3.5-122B-A10B on a single Spark node, achieving up to 51 tokens per second with vLLM patches and custom quick-start configurations. The deployment thread documents hands-on local inference performance for the MoE model on GB10 hardware. Thread participants compare vLLM 0.19.1-stable against 0.20.1 builds specifically for this model. Testing showed approximately a 9% performance regression when moving from vLLM 0.19.1 to 0.20.1 on Qwen3.5-122B-A10B, with Q&A dropping 8.3%, code 9.6%, and long-context 10.5%. Users noted vLLM 0.20.1 stabilized tool calls for stock templates but degraded MTP-enabled tool quality. Maintainers recommended staying on 0.19.1-stable pending further regression benchmarks.

An NVIDIA DGX Spark community thread by user Albond documents optimization work for Qwen3.5-122B-A10B, achieving up to 51 tokens per second on a single Spark node. The stack combines vLLM 0.19 with Intel AutoRound INT4, FlashInfer, hybrid INT4+FP8 quantization for shared expert dense layers using Qwen's official FP8 checkpoint, and MTP-1 speculative decoding reaching 95% acceptance rate. The thread notes that vLLM's default INC quantization config silently dispatched FP8 shared-expert layers through UnquantizedLinearMethod, effectively a bug fixed by a 95-line patch enabling CUTLASS FP8 execution. MTP head weights from Intel AutoRound's model_extra_tensors.safetensors (4.8 GB) can be re-enabled for vanilla checkpoints via --speculative-config, while hybrid checkpoints require the add-mtp-weights.py script to restore MTP mappings.

Alibaba's Qwen team released the Qwen3.5 Medium series on February 25, 2026, with Qwen3.5-122B-A10B as one of three open-weight models available under the Apache 2.0 license on Hugging Face and ModelScope. The model uses a hybrid architecture combining Gated Delta Networks with a sparse Mixture-of-Experts system, activating roughly 10 billion parameters per token from 122 billion total, and supports multimodal inputs alongside agentic tool calling. According to VentureBeat's launch coverage, Qwen3.5-122B-A10B delivers performance comparable to proprietary models from OpenAI and Anthropic, reportedly beating GPT-5-mini and Claude Sonnet 4.5 on third-party benchmarks even after aggressive quantization. The release emphasizes that frontier-level context windows and strong accuracy are preserved at 4-bit weight and KV cache quantization, enabling deployment on consumer-grade GPUs without server-grade infrastructure.

The Kilo Code benchmark page describes Qwen3.5-122B-A10B as a hybrid architecture combining a linear attention mechanism with a sparse mixture-of-experts model for higher inference efficiency. The native vision-language model supports a 262,144-token context window with up to 65,536 tokens of max output and multimodal input support. The page positions it for integration through the Kilo Code open-source coding agent. Enkrypt AI red-team evaluations report an overall risk score of 20.5/100 for Qwen3.5-122B-A10B, with a composite safety score of 28.2/100. Category breakdowns include bias at 59.7/100, CBRN at 29.8/100, and insecure code at 7.1/100, while harmful content measured 2.2/100 and toxicity 3.6/100. NIST and OWASP-weighted averages landed at 20.0 and 23.0 respectively, giving developers a third-party safety baseline.

According to the LLM Stats composite dashboard, Qwen3.5-122B-A10B ranks 97 overall and scores 35.0 against a blended price of $0.39 per million tokens. It places in the top 10% for healthcare, legal, and finance, while scoring below average on websites, games, tool calling, and 3D tasks. Conversation-depth analysis shows quality holds between 13.0 and 14.9 across turn 1 through turn 31+. On dataset-level benchmarks sourced from qwen.ai, Qwen3.5-122B-A10B achieves 0.97 on CountBench and 0.97 on VLMsAreBlind, with MMLU-Redux also listed. The quality tracker reports a stable +0.80σ trend over 177 seven-day votes, indicating consistent performance. Capability-tier rankings span Chat, Coding, Reasoning, and Vision categories against hundreds of competing models.

Qwen3.5-122B-A10B is a 122-billion-parameter multimodal Mixture-of-Experts model from Alibaba's Qwen team that activates roughly 10 billion parameters per token, balancing large-scale reasoning with efficient inference. The unified multimodal framework handles text and images for document understanding, chart interpretation, and visual question answering. It supports a native 256,000-token context, extendable via YaRN scaling. Released under the Apache 2.0 license, the open-weight model builds on earlier Qwen multimodal systems for developers tackling demanding multimodal reasoning. Roboflow Playground's vision benchmarks place it at 76.12% pass rate across 67 tasks, ranking 9 of 77 visual-understanding models. Per-task cost is reported at $0.0003, with token pricing at $0.26 per million input.

Ollama's library page for qwen3.5:122b-a10b distributes the model under Apache License 2.0 as an 81GB Q4_K_M quantized package built on the qwen35moe architecture with approximately 125 billion parameters. The model card confirms multimodal vision support alongside tool calling and thinking capabilities, reflecting Qwen3.5's unified vision-language foundation designed for cross-modal reasoning, coding, and agent workflows. The Ollama listing documents Qwen3.5's core architectural enhancements, including Gated Delta Networks paired with sparse Mixture-of-Experts for high-throughput inference and early-fusion multimodal training. However, the benchmark table displayed on the page reports results for the larger Qwen3.5-397B-A17B sibling rather than the 122B-A10B variant, meaning direct performance numbers for this subject should not be cited from that table.