{"slug": "qwen3-6-35b-a3b", "title": "Qwen3.6 35B A3B", "summary": "Alibaba releases Qwen3.6 35B A3B, a Mixture-of-Experts model with 35B total parameters and 3B active per token, featuring 256K native context and ranking #1 for agent workflows in 2026. The model achieves perfect tool-calling and zero catastrophic failures on a comprehensive agent benchmark, and is available under Apache 2.0 license.", "body_md": "# Qwen3.6 35B A3B\n\nMoE enthusiast35B total, 3B active per token (MoE: 256 experts, 8 routed + 1 shared), hybrid Gated DeltaNet + Gated Attention layers. 256K native context, extensible to ~1M. The agentic flagship - ranked #1 local model for agent workflows in 2026 (perfect tool-calling, zero catastrophic failures on an 84-scenario / 16-category / 8-trial benchmark). Q4_K_M (~20GB) for GGUF runtimes; Q8_0 (~37GB) for full precision.\n\n- 35.0B\n- 262k\n- apache 2.0\n- Apr 2026\n\n## Scores\n\n## Run it locally\n\nPer-quant memory needs and a static \"can you run it?\" reference - no rig entry required\n\n### Can you run it? - reference rigs\n\n| Rig | Q4_K_M | Q8_0 |\n|---|---|---|\n| NVIDIA Jetson Orin NX 16GB |\n|\n\n[no -> cloud](#cloud-pricing)Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.\n\n**How can a 24GB GPU run a 744B model?** It does not load the model into VRAM. The quantized weights (e.g. ~410GB at Q4) sit in system RAM; the GPU only holds the small shared attention and router tensors and accelerates prompt processing. Because GLM 5.2 is a Mixture-of-Experts model, each token activates only ~40B of its 744B params, so llama.cpp streams just those active experts from system RAM to the GPU each token (the `-cmoe`\n\noffload path).\n\nThat makes **decode speed bound by system-RAM bandwidth, not GPU bandwidth** - single digits on DDR4, which is why these rigs show 3-8 t/s even though they “fit.” A bigger GPU (e.g. 2x 3090) keeps more experts resident on-card and raises tok/s; a smaller GPU still runs it but pays the bandwidth tax. A 744B dense model could not run this way - only MoE’s small-active-params trick makes it possible.\n\nAggressive quants (1-2 bit) trade accuracy for size - roughly 17% accuracy loss at 2-bit vs full precision, and real long-context work often needs Q5 or Q6 even when lower quants “fit.”\n\nFormula estimates here are conservative; real tuned setups can exceed them (one HN user reports ~6 tok/s on a 512GB DDR4 + 2x 3090 rig).\n\n## Download options\n\n## Inference cost over time\n\nData accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.", "url": "https://wpnews.pro/news/qwen3-6-35b-a3b", "canonical_source": "https://tokenstead.ai/models/qwen3-6-35b-a3b", "published_at": "2026-07-09 14:12:02+00:00", "updated_at": "2026-07-09 15:08:46.656589+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-products", "ai-research"], "entities": ["Alibaba", "Qwen3.6 35B A3B", "NVIDIA Jetson Orin NX"], "alternates": {"html": "https://wpnews.pro/news/qwen3-6-35b-a3b", "markdown": "https://wpnews.pro/news/qwen3-6-35b-a3b.md", "text": "https://wpnews.pro/news/qwen3-6-35b-a3b.txt", "jsonld": "https://wpnews.pro/news/qwen3-6-35b-a3b.jsonld"}}