Qwen3.5 35B-A3B

An NVIDIA DGX Spark forum thread reports a steep performance cliff for Qwen3.5-35B-A3B as context grows: prompt processing drops from 3,918 tok/s at 2K to 60.66 tok/s at 100K, roughly a 64x slowdown. Text generation falls from 98.15 tok/s to 4.02 tok/s at 100K, making long-context agentic use feel frozen. The poster notes the model runs on the GB10 platform and the issue is observed in the Spark Arena leaderboard. The thread suggests replacing the internal MTP head with an LSTM-based speculative decoding architecture such as OWL/LongSpec, arguing it is context-agnostic and avoids the long-context degradation seen with MTP or Eagle-3 heads. The poster targets 4-agent deployments where context grows from 10K toward 150K tokens. No confirmed implementation or benchmark results are published yet.

Alibaba's Qwen team released the Qwen3.5 Medium series, including Qwen3.5-35B-A3B, as an open-weight model under Apache 2.0 on February 25, 2026, alongside the 122B-A10B and 27B variants. The 35B-A3B uses a hybrid Gated Delta Networks + sparse MoE architecture and reportedly matches Claude Sonnet 4.5 and beats GPT-5-mini on third-party benchmarks. It supports agentic tool calling and is available on Hugging Face and ModelScope. The flagship 35B-A3B can exceed 1 million token context length on consumer GPUs with 32GB of VRAM, enabled by near-lossless 4-bit weight and KV cache quantization. This brings frontier-level long context to local machines while activating only 3B parameters from its 35B total. The series was designed to remain accurate even when quantized down for desktop deployment.

Millstone's benchmark page characterizes Qwen3.5-35B-A3B-FP8 as a 35B MoE model with 256 experts (8 routed + 1 shared per pass), 3B activated parameters, and 40 transformer layers, using a Gated Delta Networks hybrid architecture. It supports a native 256K context length extendable to 1M via YaRN, and handles 201 languages plus image and video inputs. Fine-grained FP8 quantization with a 128 block size delivers near-identical accuracy to the original. Tested throughput includes 598 tok/s on a single RTX Pro 6000 Blackwell, 908 tok/s on an H100 SXM, 1,164 tok/s on dual RTX Pro 6000, and 1,479 tok/s on an H200, across concurrency 1–15 and 1K–256K context windows. Capacity planning covers code completion, chat, document processing, and agentic coding use cases.

Qwen3.5-35B-A3B is described as an Alibaba Cloud multimodal foundation model released February 2026, with 35B total parameters and 3B activated via a 256-expert Mixture-of-Experts architecture. It targets a 262K native context window (extensible to 1M) and uses grouped-query attention (16Q/2KV, head dim 256, RoPE θ=10M) with a sparse MoE FFN routing 9 of 256 experts per token. Auxiliary parameters are listed at 450M. Benchmark scores include MMLU-Pro 0.853, GPQA Diamond 0.842, SWE-bench Verified 69.2%, and Terminal-Bench 2.0 40.5%. VRAM requirements at FP16 reach 66.43 GB at 1K context, with recommended single-GPU setups including 1x RTX PRO 5000 Blackwell (72GB) or 1x A100 (80GB). The spec sheet lists API pricing at $0.25 input and $2.00 output per million tokens from a single provider reference.