Hardware Corner Mar 2026: RTX 6000 Ada Qwen3 32B Q4_K decode @ 64k ctx
11.8 tok/s (decode, per-request, batch 1, KV depth 64k)
Fields
id
benchmarks/hc26-rtx6000ada-qwen3-32b-q4k-decode-64kMetric
decode_tok_s
Value
11.8
Unit
tok/s (decode, per-request, batch 1, KV depth 64k)
Model
Qwen3 32B
Methodology
THIRD-PARTY LAB MEASURED (Hardware Corner per-GPU LLM benchmark page, 'Updated Benchmarks: March 2026'). llama.cpp, single GPU, GGUF Q4_K (page labels the quant Q4_K; nearest repo format_id is gguf-q4-k-m — Unsloth Q4_K_XL is what the older Dec-2025 ranking sweep used, and these March page numbers differ from that sweep by <0.1 tok/s on the shared 14B/16k cell, so they are the same harness, not a re-quant). Context length is the KV depth the table column names (4k/16k/64k/128k). Token Generation column = batch-1 per-request decode tok/s, NOT aggregate. Prompt Processing column = prefill tok/s at that context depth. Engine build hash, CUDA version, and exact prompt/output token counts beyond the context-depth bucket are not published on the page.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Single RTX 6000 Ada, weights fully in the 48 GB GDDR6. 70B and the long-context cells are not in the Dec 2025 ranking ingest. format_id gguf-q4-k-m is the nearest standard 4-bit GGUF; the page says Q4_K. Distinct from xd-nvidia-rtx-6000-ada-70b (llama-bench pp1024/tg1024, different harness and context).
Notes
Single RTX 6000 Ada, weights fully in the 48 GB GDDR6. 70B and the long-context cells are not in the Dec 2025 ranking ingest. format_id gguf-q4-k-m is the nearest standard 4-bit GGUF; the page says Q4_K. Distinct from xd-nvidia-rtx-6000-ada-70b (llama-bench pp1024/tg1024, different harness and context).