A community benchmark ran Qwen3.6-35B-A3B UD-Q4_K_XL at a median 18.045 generated tokens/s on a 2020-era laptop with a Intel Core i5-10500H (6 cores, 2.50 GHz), 32 GB DDR4-2667 and NVIDIA GeForce GTX 1650 Max-Q (4 GB). The test kept MoE experts in system RAM while attention and the KV cache ran on the GPU. That is measured evidence for this exact setup—not a claim that RAM capacity alone makes another computer compatible.
Independent results from 20–28 September 2026; not Local AI Stats tests.
This page's catalog file is UD-Q4_K_M, not the measured UD-Q4_K_XL file.
Generation
18.045 tokens/s median
One llama.cpp build and one laptop; no cross-hardware speed claim.
Six-task core suite
88.7% median across 5 passes
Custom task suite, not MMLU-Pro or a general quality score.
Safety checks
0/6 planted-key leaks · 0% fake-quote rate
Small repeated test on the benchmark's private-note workflow.
Runtime conditions: MoE experts in system RAM with --n-cpu-moe 999 --fit on; attention and KV cache on the GPU; flash attention on; Q8_0 KV cache; 32K context; six threads; CPU turbo off; temperature 0.2.
Limits: This is one independently published test on one machine. Core scores mix six custom tasks and some prompts are withheld for privacy. The blind panel used LLM judges, not human experts. The measured UD-Q4_K_XL file is not the catalog-selected UD-Q4_K_M file, so its speed and task scores must not be transferred to the catalog quant or to other computers.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
MMLU-Pro
85.2 / 100
Knowledge and reasoning across many subjects. Published accuracy; different evaluation settings can change the result.
Benchmark efficiency: 3.849 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: MMLU-Pro.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Solving coding problems. Version 6 only; published pass rates can vary with generation settings.
Benchmark efficiency: 3.632 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: LiveCodeBench v6.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Competition math from AIME 2026. Published answer accuracy.
Benchmark efficiency: 4.188 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: AIME26.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Benchmark efficiency: 3.777 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: HMMT Feb 26.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Fixing real software issues with tools. Agent setup and external tools affect these scores. File size covers model weights only.
Benchmark efficiency: 3.316 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: SWE-bench Verified.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
These cards say GPQA without naming the variant. Kept separate from confirmed GPQA Diamond results.
Benchmark efficiency: 3.885 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: GPQA.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Revised MMLU knowledge test. Kept separate from original MMLU and MMLU-Pro.
Benchmark efficiency: 4.215 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-reported results for this exact source model. Evaluation settings differ across model cards; the selected GGUF has not been independently retested. Table row: MMLU-Redux.
Published point estimate, not an independent head-to-head evaluation. Quantization and inference settings can change results.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 22.135 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 22.13 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-02. Apple US launch price for the base M5 Max Mac Studio, published 22 September 2026: USD 2,499. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Confirm model architecture and quantization support in your runner. The selected weights use 22.135 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.