Piccolo 27B (agentscope-ai)v1.7 · 10 s rule

Composite (secondary)

73.4 · Not ranked on this board

Cost and measurement conditions

$0.037 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,059,753 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0365791641698085/1,000; old list-price estimate USD 0.138785651443922/1,000 (qwen/qwen3.5-27b at USD 0.195/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0365791641698085/1,000. Estimated, not a bill. USD 0.0365791642 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.036480179 over all 1,500 items).

p50 latency: 0.6 s

Self-hosted latency estimate under the shared method (assumption, not measured)

Measured 30 Oct 2026 on an evaluator-owned Verda VM in FIN-03 (Finland, EU), one NVIDIA RTX PRO 6000 Blackwell Server Edition, nominal 96 GB, image 24.04.cuda13.0. THE AUTHOR'S OWN NATIVE COMPILER AND LAYOUT, measured at agentscope-ai/Piccolo-27B@9cab0ecef7372d29ac83b50ac5e49cead2caa173: piccolo_inference.py sha256 9d7253bba12525788481386da0475811db5c83e33c979551b2bc5805015b7415, piccolo_layout.py sha256 2dda6a8fd9140a5ae573950a04578de2241b8bf533247cda734b45bdeca7911f. No author HTTP server exists. Our first-party resident HTTP/usage shim loads the complete pinned safetensors checkpoint, including indexed auxiliary/MTP files, and invokes the pinned FlexCompiler and layout; the GGUF siblings were not used. The whole request is compiled before any forward. Native flex_twice defaults are preserved: max_context 65,536, row ceiling 65,535, request-wide single-read input <=3,000 tokens repeats the user input within each row when all repeated rows fit. Repetition concatenates tokens within one forward, not two independent forwards. One complete decoder call per question, use_cache=False, no chunking, truncation, prompt splitting or raised ceiling. Full-attention pre-hooks replace attention masks with dense BF16 zeros; linear layers remain unchanged. Candidate output-embedding rows are multiplied by the final hidden state in FP32, softmax temperature 1. Native choice, boolean noul and ordered score mappings are preserved; score keys are numeric strings. Input usage sums actual input_ids lengths across all question forwards, including both copies; output usage is zero, with no text generation. An over-ceiling row returns HTTP 422 and usage 0/0 for the whole request before inference. The first-party tokenizer-only counter is a separate transcription; pod-side public parity compared complete token IDs, candidates, keys, metadata and repetition behavior against the native compiler. No author code was executed on the workshop servers. The frozen TypeSafe adapter and serial driver retain the 120 s transport timeout; no host retry, answer repair, ensemble or stock-base substitution. Container base python:3.12.12-bookworm@sha256:6760942332f453ec141ea470fa3bfb790e84e11edcf0b6331b98bee6091a7b69, hash-locked upstream wheels: torch 2.14.1, CUDA 13.0, transformers 5.12.1, tokenizers 0.22.2, accelerate 1.15.0, safetensors 0.8.0. FLA, fla-core, causal-conv1d and kernels distributions are absent; Qwen3.5 linear attention and convolution use the native Torch fallback. Runtime, author-source and tokenizer hashes were retained. Source review found no benchmark-item lookup or item-specific routing; the fixed JEVLAB sentinel is a chat-template split marker. No author main() or diagnostic output mode is invoked. Author imports and numerical execution occurred only on the isolated disposable rented pod, with local hash-verified weights, read-only mounts, empty credentials, offline flags and network cut. Public parity compared complete token IDs, candidates, keys, metadata and repetition behavior with the native compiler; independent golden fixtures passed. Fresh timed probes followed separate warm-ups. Noul, choice, score, 255-choice, Unicode JSON, multi-question and repetition-boundary probes passed; overflow returned HTTP 422 with zero usage. These are synthetic observations, not scored percentiles or full-ceiling capacity proof. Separate public pre-warm passed. The complete 1,500-request run took 398.85788083076477 s; raw output sha256 864f8d24a20f52515d3d53ac4eaa981745f3d2c529930ee626f72dd983a08a1f was verified. Post-run health was ready. The VM was closed and temporary node and local copies deleted, with no cleanup errors reported. GPU rental spend is separate from the benchmark's model-cost estimate.