z3-mini-soup-preview (Zils AI)v1.7 · 10 s rule

Composite (secondary)

79.0 · Not ranked on this board

Cost and measurement conditions

$0.0037 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: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 567,051 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00374212007096435/1,000; old list-price estimate USD 0.0114248018804567/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00374212007096435/1,000. Estimated, not a bill. Reference deepinfra:Qwen/Qwen3.5-4B is frozen at the 25 Sep cut-off and was deprecated afterwards (replaced by Qwen/Qwen3.5-9B); re-score if it changes (I-3). USD 0.00374212007 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0063155701 over all 1,500 items).

p50 latency: 0.2 s

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

Measured 31 Oct 2026 on an evaluator-owned rented Verda VM in FIN-03 (Finland, EU), one NVIDIA RTX PRO 6000 Blackwell Server Edition, nominal 96 GB (97,887 MiB reported), PAYG non-spot instance type 1RTXPRO6000.30V, image 24.04.cuda13.0, driver 580.178.04. This is the measured device, without establishing a minimum. The author's native d3_server.py, d3_runtime.py and d3_format.py at model revision 1492185bb48347db2a9d88f8bf18cb0e9c17a55b ran unmodified in a pinned dependency container, with read-only source, model and input, private output, empty credentials and offline flags. The first-party launcher set D3_FAST=0 and checked NVIDIA CUDA without HIP, model.fast is None, inactive fast_report and no imported d3_kernels. Alternative modeling_d3, pipeline_d3 and d3_engine paths were not used. Native defaults: batch size 8, BF16 backbone, SDPA, noncausal full attention, float32 255-way readout, temperature 1.0, permutation averaging off and normal warm-up. One question per frozen adapter request gives one unpadded forward pass, without generated text, inference ensemble, host retry or answer repair. The shipped tokenizer/chat template uses thinking disabled, add_generation_prompt=True and add_special_tokens=False; state JSON serialization and option order are preserved. decision_config.json has max_length null, without cap, truncation or splitting. Native usage counts rendered input, with output_tokens zero. Runtime: torch 2.13.0+cu130, transformers 5.17.0, tokenizers 0.23.2, safetensors 0.8.0, accelerate 1.15.0 and hash-pinned fla-core/flash-linear-attention 0.5.2 overlay wheels. Base dependency image vllm/vllm-openai:v0.31.0@sha256:a4a4c0437bf7240089da5f08aa370c4aee17ae5290f7a3b468825ee26c4c3a6b; derived image sha256:c0d0cc0aa868be27d9502d8a1fff9537d378b530588a702b3f492991101c8097. The image supplies dependencies; the vLLM engine is not used and stock-vLLM equivalence is not claimed. All 31 pinned snapshot files, including five selected author-source files and tokenizer hashes, were verified on the measuring VM. Weight body: 4,539,265,536 BF16 parameters plus separate readout. Author source and weights ran only on the disposable rented VM, read-only, offline and without infrastructure credentials. No benchmark-item-specific branch or answer lookup was found in the reviewed serving closure. Dormant media-loading capabilities were confined by text-only requests, local-only loading and network isolation, proven before sealed input. Public native probes from 256 through 81,965 tokens and frozen-adapter probes covered the documented types, 255 choices, long state and Unicode JSON, with native usage equality. Unsupported 11-level score returned native HTTP 422 with healthy recovery. The API does not expose token IDs; no token-ID concordance is claimed. Separate public warm-ups and scored-server pre-warm passed. Complete 1,500-request run, driver exit 0, scored loop 110.09967112541199 s. Raw output sha256 65a4e01f340356f1e2d081e8124dee3ae860a18efd3089420a6293516b3e7848 was verified. Post-run health was ready. The VM was closed and temporary node and local copies deleted; provider deletion and guard release were confirmed before deadline. GPU rental spend is separate from the benchmark model-cost estimate.