MetaDecision-31B-Rainv1.7 · 10 s rule

Composite (secondary)

73.8 · Not ranked on this board

Cost and measurement conditions

$0.032 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 gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 717,572 input + 72,041 output tokens over 1,489 decisions = 530.29751511 tokens/decision. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0317219617003379/1,000; old list-price estimate USD 0.0598223102753526/1,000 (google/gemma-4-31b-it via frozen pricing_v15 reference at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0317219617003379/1,000. Estimated, not a bill. USD 0.0317219617 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0329463697 over all 1,500 items).

p50 latency: 0.7 s

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

Measured on one Verda RTX PRO 6000 Blackwell Server Edition GPU, 96 GB, FIN-03, the author's own server with network access cut. The vLLM image was vllm/vllm-openai@sha256:a4a4c0437bf7240089da5f08aa370c4aee17ae5290f7a3b468825ee26c4c3a6b; scored elapsed time was 2,422 seconds. Serving code is MeTask-Rain/MetaDecision-31B-Rain@068cb1420d40b06e48895f2fc75f29e6da37748b; weights are MeTask-Rain/metask-jev-rain-31B-trained@2feff076f07711fe2fefe7a6fdcedfe4132b773e. The weights are the same as the live rows metask_jev_rain_31b_trained and metask_jev_rain_31b_trained_v3, bound by SHA-256 (model.safetensors 1b9b1d622a93f02c0d33f98e502f233b5d707443af6ddc464ed0bf5498506c20; adapter_model.safetensors 2b9568998ffed5d133223ffdff03056be297285c643122db4d45365d0bd86970); only service/ changed. The ordered repo is the serving-code repo. Engine-reported token usage includes drafting and readout passes: 753,547 input and 72,602 output tokens across 1,500 decisions; no usage-estimated rows. The initial-prompt count was 597,680 processed tokens versus 964,173 untrimmed, with 5 trimmed items. The service extracts query-relevant state chunks, protects explicit rule sections, and marks omissions; inputs above 6,653 tokens are answered from a shortened context, and a required hard cut retains 60% of the head and 40% of the tail. A 50-item, two-pass determinism check had 1/50 argmax flip. Drafting has a time-based 9-second per-decision deadline, so answers can depend on timing. The served base is Google's W4A16 4-bit QAT checkpoint (google/gemma-4-31B-it-qat-w4a16-ct) with the unmerged rank-16 LoRA adapter applied at serving time. The service keeps state across decisions: drafting passes are capped at 49% of all decisions so far, and from the 101st decision a cost fuse limits drafting when the running engine cost nears USD 0.0646 per 1,000 decisions, priced at USD 0.0598 per million tokens (this benchmark's GPU-time rate for the architecture). Answers and cost can therefore depend on the order of decisions.

Model and serving disclosure

Re-submission on the same SHA-256-bound weights as the live metask_jev_rain_31b_trained and metask_jev_rain_31b_trained_v3 rows; only the serving code in service/ changed. Two inherited serving settings were chosen using the 231 public JevBench items: per-type temperatures in the first calibration step and the 840-token drafting-prompt limit. These choices are in-sample on the public split; the customer also used its own development data. A 50-item, two-pass determinism check had 1/50 argmax flip. The service uses a time-based 9-second drafting deadline, so answers can depend on timing. Measured on one Verda RTX PRO 6000 Blackwell 96 GB (FIN-03), the author’s own server with network access cut, using vllm/vllm-openai@sha256:a4a4c0437bf7240089da5f08aa370c4aee17ae5290f7a3b468825ee26c4c3a6b; scored elapsed time was 2,422 seconds. Engine-reported usage includes drafting/readout passes. Initial prompts: 597,680 processed tokens from 964,173 untrimmed; 5 trimmed items. The service extracts query-relevant state chunks, protects explicit rule sections, marks omissions, and uses a 6,653-token input cap; a required hard cut keeps 60% head and 40% tail. The served base is Google's W4A16 4-bit QAT checkpoint (google/gemma-4-31B-it-qat-w4a16-ct) with the unmerged rank-16 LoRA adapter applied at serving time. The service keeps state across decisions: drafting passes are capped at 49% of all decisions so far, and from the 101st decision a cost fuse limits drafting when the running engine cost nears USD 0.0646 per 1,000 decisions, priced at USD 0.0598 per million tokens (this benchmark's GPU-time rate for the architecture). Answers and cost can therefore depend on the order of decisions.