Can I run Qwen-AgentWorld-35B-A3B on a M5 Ultra 96GB Mac?
Computed with the open FitLLM engine — accurate per-layer KV-cache modeling, not a naive estimate. Updated 2026-10-19.
Conditions: 8-bit · KV cache F16 · 32K context · macOS reserve (8.0 GB).
Memory breakdown (8-bit, F16 KV, 32K context)
| Model weights | 32.3 GB |
| KV cache | 0.6 GB |
| Linear-attention state (fixed — does not grow with context) | 0.1 GB |
| Runtime + macOS | 13.0 GB |
| Total used | 46.0 / 96 GB |
| Free | 50.0 GB |
Context at 8-bit: recommended max ~256K tokens (keeps 20% headroom) · hard limit ~256K tokens. Unified memory is shared with macOS and your apps — the recommended max leaves 20% of it free.
Every quantization on M5 Ultra 96GB
| Quant | Weights | Recommended max ctx | Hard limit | Used @32K |
|---|---|---|---|---|
| 4bit | 16.2 GB | ~256K | ~256K | ✅ 27.9 / 96 GB |
| 8bit | 32.3 GB | ~256K | ~256K | ✅ 46.0 / 96 GB |
| 16bit | 64.6 GB | — | ~256K | ⚠️ 82.2 / 96 GB |
Recommended max = the longest context that still keeps the 20% headroom macOS needs · hard limit = the longest that fits at all (KV cache F16).
Lower quants free memory at some output-quality cost — 4-bit is the common sweet spot for local use.
▶ Open the interactive calculator (this exact setup)Embed this verdict
Live badge for your README or model card — recomputed by the engine at exactly these conditions, never stale:
[](https://fitllm.run/can-i-run/qwen-agentworld-35b-a3b-on-m5-ultra-96gb)
Or from your terminal (exit 0/1 — works as a pre-download guard):
npx fitllm "Qwen-AgentWorld-35B-A3B" --mac 96 --quant 8 --ctx 32768 --kv 16Why most calculators get this wrong
Qwen-AgentWorld-35B-A3B is a hybrid model: only 10 of its 40 layers use full attention — the rest are linear and keep no growing KV cache. Naive calculators count every layer at full context and badly over-estimate.
Other options
same Mac Models that fit in 96GB: GLM-4.7-Flash, gpt-oss-20b, Qwen 3.6 27B, Qwen 3.6 35B-A3B, Qwen-AgentWorld-35B-A3B, Gemma 4 e2b, Gemma 4 e4b, Gemma 4 12b, Gemma 4 26b A4B, Gemma 4 31b, Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, MiniCPM5-1B, Qwen3-0.6B, Qwen3-1.7B, Llama-3.2-1B-Instruct, Gemma-3-1B-it, Qwen 3.8 27B, Laguna XS 2.1, Spark-X2.5-4B, Granite-4.2-30B, MiniCPM5-2B, Nex-N2.5-mini.
Reproduce it
Open math: fitllm-engine (MIT), from official config.json.
All numbers are computed by the open-source fitllm-engine (MIT) from official model config.json values — reproduce or audit them yourself. Estimates; real usage varies with runtime (llama.cpp / MLX / Ollama), driver and display. Found a mismatch? Report it. · FitLLM home