CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1777.8 tok/s is the aggregate decode rate across all 256 streams, about 6.9 tok/s per stream, and the 6008 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1776.5, 1778, 1777.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q35g-i-tp4-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q35g-i-tp4-m4 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 108.6, 112.9, 105.7 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 6369.5 tok/s is the aggregate decode rate across all 256 streams, about 24.9 tok/s per stream, and the 1355 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 6381, 6356.6, 6369.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 6336.8 tok/s is the aggregate decode rate across all 192 streams, about 33.0 tok/s per stream, and the 675 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 independent one-card vLLM servers (one per card; identical but for card and port) behind a round-robin nginx router; the command shown is the first server's; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 6206.5, 6378.7, 6336.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp1r3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config lfm8-u-tp2-m0 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 165.7, 165.7, 165.7 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 3359.3 tok/s is the aggregate decode rate across all 128 streams, about 26.2 tok/s per stream, and the 1053 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 3359.3, 3374.9, 3354.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config lfm8-u-tp1-m0 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 128.9, 128.8, 128.9 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 525.7 tok/s is the aggregate decode rate across all 256 streams, about 2.1 tok/s per stream, and the 13363 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 525.4, 525.8, 525.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 324 tok/s is the aggregate decode rate across all 192 streams, about 1.7 tok/s per stream, and the 51154 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 325.4, 324, 321.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 251.8 tok/s is the aggregate decode rate across all 256 streams, about 1.0 tok/s per stream, and the 95944 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 251.8, 251.3, 251.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 116.2 tok/s is the aggregate decode rate across all 128 streams, about 0.9 tok/s per stream, and the 121198 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 116.2, 116.2, 116.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1283.5 tok/s is the aggregate decode rate across all 256 streams, about 5.0 tok/s per stream, and the 5710 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1275.3, 1283.5, 1288.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 1112.9 tok/s is the aggregate decode rate across all 192 streams, about 5.8 tok/s per stream, and the 6177 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 1113.4, 929, 1112.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 733.5 tok/s is the aggregate decode rate across all 256 streams, about 2.9 tok/s per stream, and the 8833 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 732.7, 733.5, 734 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 381.8 tok/s is the aggregate decode rate across all 128 streams, about 3.0 tok/s per stream, and the 8999 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 381.8, 381.5, 381.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1052 tok/s is the aggregate decode rate across all 256 streams, about 4.1 tok/s per stream, and the 6122 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1046.6, 1052, 1053.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 497 tok/s is the aggregate decode rate across all 192 streams, about 2.6 tok/s per stream, and the 27921 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 497, 497.1, 496.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 605.3 tok/s is the aggregate decode rate across all 256 streams, about 2.4 tok/s per stream, and the 9488 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 605, 605.5, 605.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 180.7 tok/s is the aggregate decode rate across all 128 streams, about 1.4 tok/s per stream, and the 73735 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 180.9, 180.7, 180.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1295.2 tok/s is the aggregate decode rate across all 256 streams, about 5.1 tok/s per stream, and the 12639 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1296.1, 1295.2, 1295 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 929 tok/s is the aggregate decode rate across all 192 streams, about 4.8 tok/s per stream, and the 11981 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 928.3, 929.1, 929 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 740.5 tok/s is the aggregate decode rate across all 256 streams, about 2.9 tok/s per stream, and the 20371 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 740.6, 740.4, 740.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 357.3 tok/s is the aggregate decode rate across all 128 streams, about 2.8 tok/s per stream, and the 28691 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 356.4, 358.5, 357.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1234.8 tok/s is the aggregate decode rate across all 256 streams, about 4.8 tok/s per stream, and the 9669 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (296 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1234.8, 1241.2, 1228.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config qc30-i-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 2677.4 tok/s is the aggregate decode rate across all 256 streams, about 10.5 tok/s per stream, and the 4113 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (296 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 2685.1, 2677.4, 2677.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config qc30-i-tp2-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1777.8 tok/s is the aggregate decode rate across all 256 streams, about 6.9 tok/s per stream, and the 6008 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1776.5, 1778, 1777.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q35g-i-tp4-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q35g-i-tp4-m4 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 108.6, 112.9, 105.7 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 6369.5 tok/s is the aggregate decode rate across all 256 streams, about 24.9 tok/s per stream, and the 1355 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 6381, 6356.6, 6369.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 6336.8 tok/s is the aggregate decode rate across all 192 streams, about 33.0 tok/s per stream, and the 675 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 independent one-card vLLM servers (one per card; identical but for card and port) behind a round-robin nginx router; the command shown is the first server's; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 6206.5, 6378.7, 6336.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp1r3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config lfm8-u-tp2-m0 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 165.7, 165.7, 165.7 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 3359.3 tok/s is the aggregate decode rate across all 128 streams, about 26.2 tok/s per stream, and the 1053 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (283 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 3359.3, 3374.9, 3354.1 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config lfm8-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config lfm8-u-tp1-m0 served by kasa-testing scripts/bench serve.sh after its correctness gate. Canonical reasoning-v1, lmx single-shot requests after one discarded warm-up. Measured requests: 128.9, 128.8, 128.9 tok/s; this is the median (declared rule, tie to the earlier). Prompt/output samples, engineTimingsRaw and MTP counts were added from the lmx run file and exact per-request vLLM /metrics deltas, because lmx does not send them.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 525.7 tok/s is the aggregate decode rate across all 256 streams, about 2.1 tok/s per stream, and the 13363 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 525.4, 525.8, 525.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 324 tok/s is the aggregate decode rate across all 192 streams, about 1.7 tok/s per stream, and the 51154 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 325.4, 324, 321.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 251.8 tok/s is the aggregate decode rate across all 256 streams, about 1.0 tok/s per stream, and the 95944 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 251.8, 251.3, 251.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 116.2 tok/s is the aggregate decode rate across all 128 streams, about 0.9 tok/s per stream, and the 121198 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 116.2, 116.2, 116.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g31-iq4-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1283.5 tok/s is the aggregate decode rate across all 256 streams, about 5.0 tok/s per stream, and the 5710 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1275.3, 1283.5, 1288.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 1112.9 tok/s is the aggregate decode rate across all 192 streams, about 5.8 tok/s per stream, and the 6177 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 1113.4, 929, 1112.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 733.5 tok/s is the aggregate decode rate across all 256 streams, about 2.9 tok/s per stream, and the 8833 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 732.7, 733.5, 734 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 381.8 tok/s is the aggregate decode rate across all 128 streams, about 3.0 tok/s per stream, and the 8999 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 381.8, 381.5, 381.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-if8-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1052 tok/s is the aggregate decode rate across all 256 streams, about 4.1 tok/s per stream, and the 6122 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1046.6, 1052, 1053.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 497 tok/s is the aggregate decode rate across all 192 streams, about 2.6 tok/s per stream, and the 27921 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 497, 497.1, 496.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 605.3 tok/s is the aggregate decode rate across all 256 streams, about 2.4 tok/s per stream, and the 9488 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 605, 605.5, 605.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 180.7 tok/s is the aggregate decode rate across all 128 streams, about 1.4 tok/s per stream, and the 73735 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (304 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 180.9, 180.7, 180.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g12-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1295.2 tok/s is the aggregate decode rate across all 256 streams, about 5.1 tok/s per stream, and the 12639 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1296.1, 1295.2, 1295 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 192 requests sent at once (batch 192, lmx --concurrency 192); 929 tok/s is the aggregate decode rate across all 192 streams, about 4.8 tok/s per stream, and the 11981 ms time to first token is under that load. Setup: KASA B65 rig, 3 x Intel Arc Pro B65, 3 engine replicas (data parallel), 1 card each; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 64 per replica, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 192: 928.3, 929.1, 929 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp1dp3-m0-c192, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 740.5 tok/s is the aggregate decode rate across all 256 streams, about 2.9 tok/s per stream, and the 20371 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 740.6, 740.4, 740.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 128 requests sent at once (batch 128, lmx --concurrency 128); 357.3 tok/s is the aggregate decode rate across all 128 streams, about 2.8 tok/s per stream, and the 28691 ms time to first token is under that load. Setup: KASA B65 rig, 1 x Intel Arc Pro B65, one card; vLLM 0.32.0rc1.dev100+g77154ca61 (vllm/vllm-openai-xpu:nightly-77154ca6116ae34cbfe314beef1016bde9b63989); max_num_seqs 128, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (346 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 356.4, 358.5, 357.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q38d-u-tp1-m0-c128, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1234.8 tok/s is the aggregate decode rate across all 256 streams, about 4.8 tok/s per stream, and the 9669 ms time to first token is under that load. Setup: KASA B65 rig, 4 x Intel Arc Pro B65, tensor parallel across 4 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (296 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 1234.8, 1241.2, 1228.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config qc30-i-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 2677.4 tok/s is the aggregate decode rate across all 256 streams, about 10.5 tok/s per stream, and the 4113 ms time to first token is under that load. Setup: KASA B65 rig, 2 x Intel Arc Pro B65, tensor parallel across 2 cards; vLLM 0.31.0+xpu.llmscaler (intel/llm-scaler-vllm:0.31.0-b1-experimental); max_num_seqs 256, max_model_len 4096; speculative decoding off. Each request: LocalMaxxing's canonical reasoning-v1 prompt (296 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 256: 2685.1, 2677.4, 2677.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config qc30-i-tp2-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
