first rows: Ornith-1.0-9B-NVFP4 release (gate, speed-v2, depth, GGUF matrix, coherence)
Browse files- README.md +17 -0
- ornith-9b-nvfp4-2026-07-03.jsonl +29 -0
README.md
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags: [benchmarks, quantization, nvfp4, mtp, llm-serving]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# protoLabs lab-benchmarks
|
| 7 |
+
|
| 8 |
+
Every number on a protoLabsAI model card traces to a row here. Release-gate results
|
| 9 |
+
(quant vs bf16 baseline, paired task sets, outliers re-trialed x3 both sides), speed-test-v2
|
| 10 |
+
regime matrices (InferenceMAX-style: seeded random dataset, client-side TTFT/TPOT p50/p99,
|
| 11 |
+
goodput), decode-at-depth ladders, and coherence-probe verdicts.
|
| 12 |
+
|
| 13 |
+
Methodology: single-stream-only numbers are never published without load numbers;
|
| 14 |
+
spec-decode accept% is reported per workload (random-data benches understate it ~2.5x);
|
| 15 |
+
LLM-judged suites are gated behind deterministic ones. Harness: protoLabsAI/protoLab evals.
|
| 16 |
+
|
| 17 |
+
CC-BY-4.0 — cite, steal, argue. Charts: protolabs.studio/lab
|
ornith-9b-nvfp4-2026-07-03.jsonl
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "gate", "suite": "function_call", "n": 54, "metric": "pass_rate", "value": 0.96, "baseline_bf16": 0.93, "trials": 1, "grading": "deterministic"}
|
| 2 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "gate", "suite": "reasoning_v2", "n": 24, "metric": "mean_task_score", "value": 0.684, "baseline_bf16": 0.726, "trials": 1, "grading": "solver"}
|
| 3 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "gate", "suite": "code_exec_v2", "n": 8, "metric": "mean_task_score", "value": 0.405, "baseline_bf16": 0.391, "trials": 3, "grading": "execution"}
|
| 4 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "gate", "suite": "claw_paired29", "n": 29, "metric": "mean_task_score", "value": 0.791, "baseline_bf16": 0.819, "trials": "1 (+3x outliers both sides)", "grading": "judge", "note": "T28 excluded (bimodal judge); T12 reproducible miss 0/3 (duplicate detection)"}
|
| 5 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "coherence", "suite": "depth_probe", "n": 4, "metric": "rungs_clean", "value": "4/4 (4K/16K/32K/60K)", "grading": "detectors+adversarial-judge"}
|
| 6 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "chat-c1", "metric": "output_tok_s", "value": 132.2, "ttft_p50_ms": 46.8, "tpot_p50_ms": 7.69, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 7 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "chat-c8", "metric": "output_tok_s", "value": 907.1, "ttft_p50_ms": 70.5, "tpot_p50_ms": 8.01, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 8 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "context-c1", "metric": "output_tok_s", "value": 108.1, "ttft_p50_ms": 343.2, "tpot_p50_ms": 8.92, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 9 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "context-c8", "metric": "output_tok_s", "value": 598.0, "ttft_p50_ms": 412.3, "tpot_p50_ms": 12.23, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 10 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "bf16+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "chat-c1", "metric": "output_tok_s", "value": 82.7, "ttft_p50_ms": 90.1, "tpot_p50_ms": 12.4, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 11 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "bf16+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "chat-c8", "metric": "output_tok_s", "value": 620.8, "ttft_p50_ms": 121.0, "tpot_p50_ms": 12.37, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 12 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "bf16+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "context-c1", "metric": "output_tok_s", "value": 73.1, "ttft_p50_ms": 656.8, "tpot_p50_ms": 13.24, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 13 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "bf16+mtp", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "context-c8", "metric": "output_tok_s", "value": 385.9, "ttft_p50_ms": 733.3, "tpot_p50_ms": 18.45, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 14 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d16k-c1", "metric": "output_tok_s", "value": 64.1, "ttft_p50_ms": 732.0, "tpot_p50_ms": 13.28, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 15 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d16k-c4", "metric": "output_tok_s", "value": 159.7, "ttft_p50_ms": 834.8, "tpot_p50_ms": 21.64, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 16 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d32k-c1", "metric": "output_tok_s", "value": 42.1, "ttft_p50_ms": 1709.3, "tpot_p50_ms": 17.92, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 17 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d32k-c4", "metric": "output_tok_s", "value": 88.9, "ttft_p50_ms": 2265.5, "tpot_p50_ms": 36.8, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 18 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d4k-c1", "metric": "output_tok_s", "value": 104.3, "ttft_p50_ms": 139.2, "tpot_p50_ms": 9.5, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 19 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d4k-c4", "metric": "output_tok_s", "value": 311.1, "ttft_p50_ms": 191.7, "tpot_p50_ms": 11.38, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 20 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d64k-c1", "metric": "output_tok_s", "value": 24.3, "ttft_p50_ms": 4081.7, "tpot_p50_ms": 26.9, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 21 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "nvfp4+mtp-depth", "hardware": "RTX PRO 6000 Blackwell (sm120)", "engine": "vLLM 0.22.1", "regime": "d64k-c4", "metric": "output_tok_s", "value": 44.4, "ttft_p50_ms": 4631.7, "tpot_p50_ms": 71.76, "method": "vllm bench serve, random dataset seed 42, client-side"}
|
| 22 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-Q4_K_M-spec-none", "hardware": "ampere-a6000", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 104.6, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": ""}
|
| 23 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-Q4_K_M-spec-draft-mtp", "hardware": "ampere-a6000", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 153.4, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": "acceptance 0.52 (NVFP4) vs 0.49 (Q4_K_M) same-box"}
|
| 24 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-NVFP4-spec-none", "hardware": "ampere-a6000", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 70.7, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": ""}
|
| 25 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-NVFP4-spec-draft-mtp", "hardware": "ampere-a6000", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 84.8, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": "acceptance 0.52 (NVFP4) vs 0.49 (Q4_K_M) same-box"}
|
| 26 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-Q4_K_M-spec-none", "hardware": "blackwell-sm120", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 205.1, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": ""}
|
| 27 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-Q4_K_M-spec-draft-mtp", "hardware": "blackwell-sm120", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 239.0, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": "acceptance 0.52 (NVFP4) vs 0.49 (Q4_K_M) same-box"}
|
| 28 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-NVFP4-spec-none", "hardware": "blackwell-sm120", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 201.5, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": ""}
|
| 29 |
+
{"model": "deepreinforce-ai/Ornith-1.0-9B", "quant": "nvfp4-w4a4-calibrated", "artifact": "protoLabsAI/Ornith-1.0-9B-NVFP4", "date": "2026-07-03", "harness": "protoLab evals @ claw-eval 49d2b6a", "judge": "protolabs/reasoning", "kind": "speed", "config": "gguf-NVFP4-spec-draft-mtp", "hardware": "blackwell-sm120", "engine": "llama.cpp b9829", "metric": "gen_tok_s", "value": 306.0, "method": "llama-cli -n 200 greedy; blackwell mtp = mean n=6 prompts", "note": "acceptance 0.52 (NVFP4) vs 0.49 (Q4_K_M) same-box"}
|