Qwen3.6 35B-A3B

The news blog specialized in Japanese culture, odd news, gadgets and all other funny stuffs. Updated everyday.

BenchLM aggregates Qwen3.6-35B-A3B benchmark data into a capability score of 43.6/100 across 41 published rows, with category-specific rankings: Agentic rank 119 of 157 (40.6, 11 benchmarks verified), Coding rank 95 of 159 (45.3, 6 verified), Multimodal rank 39 of 50 (50.0, 14 verified), and Knowledge plus Math categor Reported operational characteristics include a measured speed of 115 tokens/second against a field median of 90.5 tok/s, a first-token latency of 48.94 seconds, and a 262K-token context window, though no first-party hosted API token rate is published. BenchLM highlights multimodal strength as the model's standout, part

The benchmarks.company model card for Qwen3.6-35B-A3B documents developer-reported scores across major suites: SWE-bench Verified at 73.4% resolved (agentic), SWE-bench Pro at 49.5% pass@1 (coding), Terminal-Bench 2.0 at 51.5% task resolution (agentic), LiveCodeBench at 80.4% pass@1, and MMLU-Pro at 85.2% accuracy. The A hardware-fit table maps usable VRAM per quantization — IQ3_XXS at 14GB, Q4_K_M at 21GB, Q8_0 at 37GB, and FP16 at 70GB — against consumer and professional GPUs ranging from the RTX 3060 12GB through the RTX PRO 6000 Blackwell 96GB and M5 Max 128GB unified memory. The page notes that the MoE's low active-parameter cou

A community benchmark on prismix.dev compared tool-calling performance of original Qwen3.6-35B-A3B against several fine-tunes — KAT-Coder, Ornith 1.5, and the newer Tiel-Coder — using the tool-eval-bench suite on a cluster of 32GB V100s. Testing 13 GGUF quantizations with five runs each (over 300 GPU-hours total), Orni The benchmark used llama.cpp build 10433 with q8_0 KV cache and ubatch-size 2048, focusing on prompt-processing-heavy workloads rather than raw speed. Results provide practical guidance for users running Qwen3.6-35B-A3B derivatives on VRAM-limited hardware around Q4 quant sizes (15–22GB GGUF files), highlighting which

Alibaba's Tongyi Lab released Qwen3.6-35B-A3B as an open model on April 15, 2026, following the April 2 announcement of Qwen3.6-Plus. The model is a sparse mixture-of-experts design with 35 billion total parameters but only 3 billion active parameters, and ships under the Apache 2.0 license. Tongyi Lab positions it as The release emphasizes efficient inference for VRAM-constrained hardware, leveraging the small active-parameter count to keep deployment costs low. The GIGAZINE report links to Tongyi Lab's official blog post and X announcements framing the model as an open-source coding-focused entry point into the Qwen3.6 family. Ind

On April 2, 2026, Alibaba's Qwen team open-sourced Qwen3.6-35B-A3B, the first open-weight variant of the Qwen3.6 generation, releasing it under Apache 2.0 on Hugging Face alongside the proprietary Qwen3.6-Plus API model. The 35-billion-parameter Mixture-of-Experts model activates just 3B parameters per token while post The open weights are compatible with Hugging Face Transformers, vLLM (≥0.19.0), SGLang (≥0.5.10), and KTransformers, with an unsloth GGUF build appearing on Hugging Face within hours of launch. Qwen3.6-35B-A3B is a sparse MoE with 256 experts (8 routed plus 1 shared per token), arranged in a 40-layer stack using a repe

OpenRouter's listing for Qwen3.6-35B-A3B confirms it is an open-weight multimodal model from Alibaba Cloud with 35 billion total parameters and 3 billion active parameters per token, built on a hybrid sparse mixture-of-experts architecture that combines Gated DeltaNet linear attention with standard gated attention laye OpenRouter routes the same weights across at least nine providers with meaningfully different price-performance profiles. Listed rates span Darkbloom at $0.05 input / $0.70 output per 1M tokens (1.27s P50 latency, 88 tps, 99.80% uptime) up to CoreWeave at $0.25 / $1.25 (0.33s, 156 tps, 99.95%), with DeepInfra, Venice,