CONCURRENCY TEST, NOT A SINGLE-USER RESULT. 256 requests sent at once (batch 256, lmx --concurrency 256); 1159.9 tok/s is the aggregate decode rate across all 256 streams, about 4.5 tok/s per stream, and the 4731 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: 1154.4, 1159.9, 1160 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g26-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); 282.9 tok/s is the aggregate decode rate across all 128 streams, about 2.2 tok/s per stream, and the 49086 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 (305 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 282.2, 282.9, 284.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g26-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); 3777 tok/s is the aggregate decode rate across all 256 streams, about 14.8 tok/s per stream, and the 2583 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: 3779.5, 3777, 3768.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36x-if8-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); 2796.8 tok/s is the aggregate decode rate across all 256 streams, about 10.9 tok/s per stream, and the 3277 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: 2807.4, 2796.8, 2785.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36x-if8-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); 860.9 tok/s is the aggregate decode rate across all 256 streams, about 3.4 tok/s per stream, and the 18920 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: 860.9, 860.9, 860.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36e-u-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); 414.4 tok/s is the aggregate decode rate across all 256 streams, about 1.6 tok/s per stream, and the 33511 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: 414.3, 414.6, 414.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36e-u-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); 1407.1 tok/s is the aggregate decode rate across all 256 streams, about 5.5 tok/s per stream, and the 9963 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: 1407.2, 1407.1, 1406.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-if8-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); 714.1 tok/s is the aggregate decode rate across all 256 streams, about 2.8 tok/s per stream, and the 16176 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: 714.3, 714.1, 714 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-if8-tp2dp1-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); 3811.2 tok/s is the aggregate decode rate across all 256 streams, about 14.9 tok/s per stream, and the 3296 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: 3811.9, 3802.3, 3811.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 3079.2 tok/s is the aggregate decode rate across all 192 streams, about 16.0 tok/s per stream, and the 3381 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: 2323.6, 3079.2, 3079.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 2609.4 tok/s is the aggregate decode rate across all 256 streams, about 10.2 tok/s per stream, and the 4972 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: 2610.8, 2609.4, 2608.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1587.5 tok/s is the aggregate decode rate across all 128 streams, about 12.4 tok/s per stream, and the 4789 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: 1585.1, 1588.5, 1587.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 3632.4 tok/s is the aggregate decode rate across all 256 streams, about 14.2 tok/s per stream, and the 3899 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: 3634.1, 3621.9, 3632.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1522.8 tok/s is the aggregate decode rate across all 192 streams, about 7.9 tok/s per stream, and the 3386 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: 1528.7, 1515.4, 1522.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 2506 tok/s is the aggregate decode rate across all 256 streams, about 9.8 tok/s per stream, and the 5006 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: 2506.6, 2506, 2505.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1496.6 tok/s is the aggregate decode rate across all 128 streams, about 11.7 tok/s per stream, and the 4798 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: 1496.6, 1496.7, 1491.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1379.8 tok/s is the aggregate decode rate across all 256 streams, about 5.4 tok/s per stream, and the 10293 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: 1379.8, 1379.1, 1379.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 525.7 tok/s is the aggregate decode rate across all 192 streams, about 2.7 tok/s per stream, and the 11902 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: 526.3, 525.7, 462.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 837.9 tok/s is the aggregate decode rate across all 256 streams, about 3.3 tok/s per stream, and the 16716 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: 838.3, 837.8, 837.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-iq4-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q38d-u-tp4-m5 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: 172.2, 162.9, 172.3 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.
KASA B65 rig. Config q38d-u-tp2-m5 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: 111.4, 105.2, 111.4 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.
KASA B65 rig. Config q38d-u-tp1-m5 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: 75.4, 74, 79.8 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); 232.3 tok/s is the aggregate decode rate across all 128 streams, about 1.8 tok/s per stream, and the 15668 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: 232.4, 232.1, 232.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 1339.9 tok/s is the aggregate decode rate across all 256 streams, about 5.2 tok/s per stream, and the 10178 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: 1339.9, 1339.7, 1339.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q34-u-tp4-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: 124.1, 124, 124 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); 1159.9 tok/s is the aggregate decode rate across all 256 streams, about 4.5 tok/s per stream, and the 4731 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: 1154.4, 1159.9, 1160 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g26-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); 282.9 tok/s is the aggregate decode rate across all 128 streams, about 2.2 tok/s per stream, and the 49086 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 (305 tokens) with a unique nonce, 256 output tokens, temperature 0. One discarded warm-up batch, then 3 measured batches of 128: 282.2, 282.9, 284.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config g26-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); 3777 tok/s is the aggregate decode rate across all 256 streams, about 14.8 tok/s per stream, and the 2583 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: 3779.5, 3777, 3768.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36x-if8-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); 2796.8 tok/s is the aggregate decode rate across all 256 streams, about 10.9 tok/s per stream, and the 3277 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: 2807.4, 2796.8, 2785.6 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36x-if8-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); 860.9 tok/s is the aggregate decode rate across all 256 streams, about 3.4 tok/s per stream, and the 18920 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: 860.9, 860.9, 860.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36e-u-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); 414.4 tok/s is the aggregate decode rate across all 256 streams, about 1.6 tok/s per stream, and the 33511 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: 414.3, 414.6, 414.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36e-u-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); 1407.1 tok/s is the aggregate decode rate across all 256 streams, about 5.5 tok/s per stream, and the 9963 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: 1407.2, 1407.1, 1406.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-if8-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); 714.1 tok/s is the aggregate decode rate across all 256 streams, about 2.8 tok/s per stream, and the 16176 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: 714.3, 714.1, 714 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-if8-tp2dp1-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); 3811.2 tok/s is the aggregate decode rate across all 256 streams, about 14.9 tok/s per stream, and the 3296 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: 3811.9, 3802.3, 3811.2 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 3079.2 tok/s is the aggregate decode rate across all 192 streams, about 16.0 tok/s per stream, and the 3381 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: 2323.6, 3079.2, 3079.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 2609.4 tok/s is the aggregate decode rate across all 256 streams, about 10.2 tok/s per stream, and the 4972 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: 2610.8, 2609.4, 2608.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1587.5 tok/s is the aggregate decode rate across all 128 streams, about 12.4 tok/s per stream, and the 4789 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: 1585.1, 1588.5, 1587.5 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 3632.4 tok/s is the aggregate decode rate across all 256 streams, about 14.2 tok/s per stream, and the 3899 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: 3634.1, 3621.9, 3632.4 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1522.8 tok/s is the aggregate decode rate across all 192 streams, about 7.9 tok/s per stream, and the 3386 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: 1528.7, 1515.4, 1522.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 2506 tok/s is the aggregate decode rate across all 256 streams, about 9.8 tok/s per stream, and the 5006 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: 2506.6, 2506, 2505.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1496.6 tok/s is the aggregate decode rate across all 128 streams, about 11.7 tok/s per stream, and the 4798 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: 1496.6, 1496.7, 1491.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q359-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); 1379.8 tok/s is the aggregate decode rate across all 256 streams, about 5.4 tok/s per stream, and the 10293 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: 1379.8, 1379.1, 1379.8 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 525.7 tok/s is the aggregate decode rate across all 192 streams, about 2.7 tok/s per stream, and the 11902 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: 526.3, 525.7, 462.7 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 837.9 tok/s is the aggregate decode rate across all 256 streams, about 3.3 tok/s per stream, and the 16716 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: 838.3, 837.8, 837.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-iq4-tp2dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q38d-u-tp4-m5 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: 172.2, 162.9, 172.3 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.
KASA B65 rig. Config q38d-u-tp2-m5 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: 111.4, 105.2, 111.4 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.
KASA B65 rig. Config q38d-u-tp1-m5 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: 75.4, 74, 79.8 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); 232.3 tok/s is the aggregate decode rate across all 128 streams, about 1.8 tok/s per stream, and the 15668 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: 232.4, 232.1, 232.3 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-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); 1339.9 tok/s is the aggregate decode rate across all 256 streams, about 5.2 tok/s per stream, and the 10178 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: 1339.9, 1339.7, 1339.9 tok/s; this is the median. Not eligible for the Verified badge, which requires batch 1. Config q36d-u-tp4dp1-m0-c256, kasa-testing campaign 2026-10-15-b65-localmaxxing-models.
KASA B65 rig. Config q34-u-tp4-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: 124.1, 124, 124 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.
