Xevn-4Bv1.7 · 10 s rule

Composite (secondary)

70.2 · Composite rank #11

Cost and measurement conditions

$0.014 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026): base string Qwen/Qwen3.5-4B priced as Qwen/Qwen3.5-4B USD 0.03/M input, USD 0.15/M output. The module selects the deprecated-after-cutoff DeepInfra reference retained under the frozen policy (price snapshots of 25 Sep 2026), USD 0.03/M input and USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Server-reported usage is supplied by our labelled first-party shim: exact native token_ids length summed over questions, with output_tokens=0; the native engine generates no text. Reported usage is 674,634 input tokens and 0 output tokens, matching the evaluator's offline count for the 1,485 answered items, offset 0 and 0 item-level mismatches. The 15 attempted long items fail with HTTP 500 CUDA OOM and report no usage; they contribute 0 reported tokens, not their 1,179,066 offline prompt tokens. Those are the same standard long items excluded from every row's common 1,485-item cost basis (v1.6.30), so common-basis usage is also 674,634 input tokens and 0 output tokens. Measured arithmetic: USD 0.013628970 per 1,000 decisions on the common basis = 0.210992220x the unchanged USD 0.06459465517241379 cap; separately USD 0.013492680/1,000 over all 1,500 items = 0.208882298x. COST ON THE COMMON BASIS: USD 0.013629 per 1,000 decisions on 1,485 items, 0.211 times the official cost cap USD 0.064595; estimated, not charged.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (97,887 MiB), FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, USD 2.5991917808219176/h recorded by the runner. Native engine and first-party HTTP shim in a --network none --read-only container; serial requests, read-only code/model/input, blank credential variables and HF offline flags. In-container cut proof before sealed input: only loopback, no default routes, DNS and TCP fail; model inventory, author and shim hashes checked. Measured 24 Oct 2026 09:20-09:26Z, about USD 0.29 rental; measurement container 126 s, run cap 1592 s, complete=true, deadline_hit=false, rc 0; scored loop 81.776480 s over 1,500 requests. Frozen v1.6-regular-20261012 self-hosted input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, checked in the container. Raw output sha256 2526d8bfedc67e8a680615052231a6c32fa62aa800b2e31cbe0aaad7afe1001b. Row cleanup and final node shred verified; exact-id VM/volume teardown confirms provider absent and guard closed at 09:26:37Z. GPU rental is not the benchmark's estimated model tariff.

Model and serving disclosure

THE AUTHOR'S OWN INFERENCE, run unmodified: quevirus/Xevn-4B at ba11f0c2d8100a1e8b7789a1174ddee7c24dcab0, sole 139-line author file xevn_infer.py (sha256 3c0e434c03e7cc96675afb86b5a9cb5f76f3af07d20f64035329600512f4a267), wrapped by our first-party shim_xevn.py for HTTP, usage and per-item error recovery. Native Xevn('/model', dtype='bfloat16', device='cuda'), eager text forward, use_cache=False, FP32 final-position hidden state and selected tied-head readout. No generation, chat template, quantization or SDPA substitution. Native calibration preserved: choice temperature 1.1, noul and score 1.2, score_lambda=0; when 0.2<P(yes)<0.8, noul snaps to 0.2 below 0.5 and to 0.8 otherwise (tie 0.5 -> 0.8). Score is the native expected ordinal. Torch 2.13.0, Transformers 5.18.0, tokenizers 0.23.2, safetensors 0.8.0, huggingface-hub 1.33.0; optional fla/causal_conv1d/kernels excluded. vLLM 0.30.0 is the image substrate, not the serving backend. Base image vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90; derived image sha256:bebbb6b15cdef220e99bca031ae8b13a0c48a622e4bfcb01afe48ca07da6f0e5. The submitted code was reviewed and cleared before the run, together with evaluator exposure, the device and native inference without an input cap, as preregistered. The runner was reviewed, findings fixed, and the final changes and hashes rechecked before the run. Source review records Hub-capable from_pretrained and conditional hf_hub_download paths; local /model with temps.json, exact inventory/hashes, offline flags and network-none avoid runtime downloads. No author shell/subprocess, eval/exec, pickle or trust_remote_code path in the reviewed serving code. Weights: Apache-2.0, public, ungated; metadata 4,659,865,088 parameters, two BF16 safetensors shards totalling 9,319,828,160 bytes. FIRST MEASUREMENT of Xevn-4B (our own intake). AUTHOR'S OWN INFERENCE on our rented hardware, one request at a time; first-party HTTP/usage shim preserves the native prompt, probabilities, temperature calibration and noul 0.2/0.8 confidence hardening. NATIVE OUTPUTS: choice argmax and probabilities, noul P(yes), score expected ordinal and probabilities. AUTHOR EXPOSURE: intake records author-reported English/Korean suite figures, not a comparable Decision Index Full or JevBench draw result; no overlap or contamination audit was run by us. The author's reported figures are not independently verified by us. CONTEXT: native, NO input cap, no evaluator-imposed cap, no truncation, splitting or fallback (ruling on file). The 16,384-token long-item boundary is a preregistration/fit classification, not a serving limit. All 1,500 items were ATTEMPTED. NOT ANSWERED, counted wrong: exactly the 15 pre-registered long positions in the pre-registered long-item set failed with HTTP 500 cuda_out_of_memory; native BF16 eager attention cannot hold these approximately 76k-80.5k-token prompts on the 96 GB card. One BF16 attention score tensor alone is approximately 186-215 GB at the 76,309-82,000-token planning/synthetic bounds. The shim frees CUDA memory after each failed item and keeps serving; no author inference change. This is the pre-registered class B complete outcome, with no failure outside the long set. Answered 1,485 of 1,500 items HTTP 200; longest normal item 1,789 native tokens. 10 s RULE: 0 decisions over 10 s; the 15 OOM items already count wrong as failures, and no additional decision is counted wrong by the time rule in the v1.7 share view (frozen v1.6 and v1.7-off views apply no time limit). Maximum 1.070494 s, median 0.041094 s, p95 0.072361 s over all 1,500 requests; OOM items fail in about 1 s. DETERMINISM: first 50 positions run twice, 50/50 bit-identical, 0 argmax flips, maximum probability difference 0.0. Pre-sealed synthetic probes: 16,384 tokens answered HTTP 200; 76,309 and 82,000 tokens attempted and failed HTTP 500, followed each time by healthy service, a successful short request and GPU memory back to about 9.5 GB.