Red Banana Lab · Field notes No. 8

Frontier MoE on a Laptop: 122 Billion Parameters, 10B Active, and the Sparse Activation Victory

Red Banana Lab, New Delhi · 14 October 2026 · production verification on our 128 GB workstation, Radeon 8060S Vulkan backend

In Field Notes No. 7 we demonstrated how a dense 32B model fits cleanly into 28 GB of resident memory, but runs into the physical wall of unified memory bandwidth: streaming 20 GB of weights per token tops out at ~8 tokens per second. Moving upward to a dense 70B model slows generation even further to ~5 tokens per second. This note documents our testing of frontier-class sparse Mixture-of-Experts architectures, specifically Qwen3.5-122B-A10B. By decoupling total parameter capacity (122B) from per-token compute volume (10B active), consumer unified memory hardware can execute 100B+ class reasoning at conversational interactive speeds.

The Architectural Comparison

Model	Architecture	Total / Active Params	Weight File	Memory Footprint
qwen2.5:32b (Instruct)	Dense Transformer	32.8B / 32.8B	19.8 GB	28.0 GB (32k ctx)
llama-3.3:70b (Projected)	Dense Transformer	70.6B / 70.6B	42.5 GB	54.0 GB (16k ctx)
gpt-oss-120b (Baseline)	Mixture of Experts	117.0B / 5.1B	61.0 GB	70.0 GB (32k ctx)
Qwen3.5-122B-A10B	Mixture of Experts	122.0B / 10.0B	42.0 GB	50.0 GB (32k ctx)

Headline Performance Measurements

Metric	Qwen2.5 32B (Dense)	Qwen3.5 122B-A10B (MoE NVMe)	gpt-oss-120b (MoE)
Generation throughput, tok/s	7.90	27.40	51.0
Weights streamed per token	19.8 GB	~3.5 GB	~2.0 GB
Prompt ingest rate, tok/s	90.9	44.1	644.0
Theoretical bus saturation	~256 GB/s (98%)	~100 GB/s (39%)	~102 GB/s (39%)
Initial cold load (NVMe mmap)	3.2 s	7.8 s	11.4 s
Initial cold load (USB 3.0)	18.5 s	49.2 s	72.0 s
Extended reasoning channel	No (Direct emit)	Yes (Native CoT)	Yes (Separate channel)
Planted verification pass	5/5	5/5	4/5

Finding 1: Sparse Activation Breaks the Memory Wall

122 Billion parameters running 3.4x faster than 32 Billion parameters On any unified memory computer, memory bus bandwidth is the ultimate governor of token generation speed. A dense 32B model must read all 19.8 GB of weights from memory for every single token produced. Qwen3.5-122B routes each token to only 10B active parameters, streaming approximately 3.5 GB of weights per token. Directly mapped on high-speed PCIe Gen4 NVMe, the 122B model delivers 27.40 tokens per second compared to 7.90 tokens per second on the 32B model: delivering nearly four times the parametric knowledge at 3.46x the interactive generation throughput.

Finding 2: The IQ3_XXS Quantization Miracle

Deploying a 122B model on a portable machine previously required 70 to 80 GB of VRAM. Using importance-matrix calibrated quantization (Unsloth Dynamic IQ3_XXS), the 122B weights occupy only 42.0 GB on disk. When fully offloaded into the Radeon 8060S Vulkan address space with a 4k context window, total allocation is approximately 50.0 GB. On our 128 GB unified hardware, this leaves over 65 GB of memory available for system buffers, browser engines, and concurrent tasks.

Finding 3: Native Chain of Thought vs Channel Budget Exhaustion

Field Notes No. 1 detailed the failure mode of gpt-oss-120b: allocating all output tokens to an internal reasoning buffer and exhausting the token budget before emitting visible text. Qwen3.5-122B resolves this structural defect. It formats its reasoning trace natively within visible thought blocks, completing step-by-step logic verification before smoothly concluding the output. In our planted multi-digit arithmetic test (48291 + 19482 + 91023), the model verified intermediate sub-sums within its thought process and returned the exact integer 158796.

Finding 4: Storage Tiering and NVMe mmap Latency

When running weights over 40 GB, the physical storage tier dictates responsiveness. Loading the 42 GB GGUF over external USB 3.0 (5 Gbps, ~90 MB/s real read) required 49.2 seconds before the first token could be processed. Offloading cold assets to our second workstation freed local high-speed PCIe Gen4 NVMe space. Placing the model on local NVMe storage cut initial kernel mmap time to 7.8 seconds, enabling near-instantaneous model launches.

Production Conclusions

For organizations deploying local intelligence on unified laptop hardware, dense models larger than 32B represent an inefficient operational choice. A 70B dense model is constrained by bus bandwidth to sub-interactive speeds. In contrast, 100B+ sparse Mixture-of-Experts architectures provide superior reasoning, deeper knowledge retention, and conversational generation rates. Red Banana Lab has standardized on 122B MoE for deep architectural analysis and 32B dense for everyday low-latency task automation.