MoE vs Dense on a 24GB Card: Qwen3.6-35B-A3B or a Dense 27B?

Short answer

Choose the MoE (Qwen3.6-35B-A3B) if your priority is:

• decode speed: only 3B of 35B params activate per token
• long interactive agent sessions where tokens/sec is felt constantly
• built-in vision, Apache 2.0

Choose a dense 27B (Qwen3.6-27B or Qwen3.8-27B) if your priority is:

• VRAM headroom: ~16-17GB at Q4 versus 20-22GB for the MoE
• longer context on the same card
• predictable quality per parameter

One rule up front: MoE saves compute, not VRAM. All experts stay resident. A 35B MoE needs 35B-class memory even though it computes like a 3B model.

The concrete numbers

The size gap is the whole argument. On a 24GB card the MoE fits, but with only 2-4GB left for KV cache — short context, no slack. The dense 27B leaves 7-8GB, which buys real context. On a 32GB card (R9700, 5090), the MoE’s squeeze disappears and its speed advantage comes free.

Speed: where MoE earns its complexity

llmcheck.net measured Qwen3.6-35B-A3B at ~58 tok/s on a 24GB Mac, versus ~24 tok/s for dense Gemma 4 31B and an estimated ~14 tok/s for a hypothetical dense 30B. Those are one site’s measurements, but the ratio matches the math: 3B active parameters cost roughly a tenth of the compute per token.

On a 4090, one benchmark site measured roughly 2x for the MoE over the dense model. Expect the gap to be larger on bandwidth-starved machines and smaller on fast GPUs.

For an agent that streams hundreds of edits per session, 58 versus 24 tok/s is the difference between waiting and not waiting. That is why speed-sensitive OpenClaw users default to the MoE — see best local models for OpenClaw.

Quality: the debate has flagships on both sides

The MoE side has receipts: one review measured Qwen3.6-35B-A3B at 73.4% on SWE-Bench while running locally — real coding-agent capability from 3B active params. And on llmcheck’s leaderboard, MoE models hold 7 of the top 10 entries.

But the dense side got a fresh flagship in August 2026: Meta chose a dense architecture for Muse Glimmer 30B precisely for agent quality on one GPU. Dense models spend every parameter on every token, and many practitioners hold that dense gives the best quality-per-param consistency — an editorial judgment, but a common one. Our Qwen3.8-27B vs Glimmer comparison covers the strongest current dense options; Laguna S-2.1 vs Qwen3.6-27B covers the previous round.

The frontier itself has voted: DeepSeek V4-Flash runs 284B total with 13B active, GLM-5.2 runs 753B with ~40B active. The 24GB debate is a microcosm of tiny-active-ratio MoE taking over the top end — while dense holds the “one GPU, maximum quality” niche.

Decision table

Final recommendation

On exactly 24GB: dense 27B first. The MoE’s 20-22GB quant leaves too little KV room for serious agent context, and the newest dense models (Qwen3.8-27B, Glimmer 30B) are strong.

If you can hold context short — or you have 32GB — Qwen3.6-35B-A3B gives you 2-4x decode speed for free. Fit first, then speed. That order never reverses.

Next steps

• Qwen3.8-27B vs Muse Glimmer 30B
• Best local LLMs for 32GB RAM
• Best local models for OpenClaw

Sources

• Qwen3.6-35B-A3B on Hugging Face
• llmcheck.net local benchmarks
• ggml-org and unsloth GGUF repos (Hugging Face)