Standard One 3B SH v4 (Standard Thinking)v1.7 · 10 s rule

Composite (secondary)

24.4 · Composite rank #96

Cost and measurement conditions

$0.052 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference, the same estimate as the live row standard_one_3b_sh: the model card names base_model StandardThinking/StandardOne-3B, built on mistralai/Ministral-3-3B-Instruct-2512. Neither Standard One checkpoint is listed on OpenRouter; the exact underlying base is priced as OpenRouter mistralai/ministral-3b-2512, list price USD 0.10/M input and USD 0.10/M output (the same OpenRouter document the live sibling cites, read 2026-10-12T00:14Z, source/public-docs/OPENROUTER-MINISTRAL-3-PRICES.json; the frozen module catalogue carries the same 0.10/0.10). Self-hosted weights, no bookable tariff of the measured deployment; the author's stated hosted list price (USD 0.038/M input, standardthinking.ai/pricing) is below this floor, so the base-model reference price applies as a floor. Note on the module: the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026) has no explicit listing decision for this base string (reference() raises), so the price is the registry value 0.10/0.10 carried over from the live entry and verified against that snapshot, not a module lookup. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the usage block of the author's own server: input_tokens = 764,986 on the 1,485 answered items, output_tokens = 0 on every row (the schema head generates no text; priced at the reference output price). Our offline count through a first-party transcription of the author's renderer and the pinned tokenizer predicted the server's numbers EXACTLY, item by item and in total (764,986 offline = 764,986 native), and equals the 12 Oct receipt of the v2 row (same tokenizer, template and item set). The 15 refused items (HTTP 422, over the author's 32,768-token context) count 0 tokens.

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), pay-as-you-go non-spot, instance type 1RTXPRO6000.30V at USD 2.519/h. The authors' reference server code/serve_head.py (127.0.0.1:30171, bf16, torch SDPA, no prefix cache) runs in ONE docker container with network none, driven one request at a time over loopback on the VM by the unchanged pinned pod_driver_v16 (frozen runtime modules sha256-pinned in pre-registration Amendment A1). Requests over the documented 32,768-token context answer HTTP 422 (refused, not truncated). The standard self-hosted x2 + 0.15 s latency adjustment applies. Network cut proven from inside the container before any item hash or model load (loopback only, no default route, DNS and TCP fail); read-only code, read-only model and input, empty credential variables (names-only proof), HF offline flags. Measured 23 Oct 2026 (VM 07:12-07:19Z; measurement 66 s, total 359 s; scored run 46 s) on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, re-checked inside the container). The input was shredded from the node and its absence verified, and the VM was removed by exact id afterwards (teardown confirmed 07:19Z).

Model and serving disclosure

THE AUTHORS' OWN REFERENCE SERVER, run unmodified, on NEW WEIGHTS: StandardThinking/StandardOne-3B-SH at 7f56f73be882bc3fb2b69abe609bab1de8297d77 (tag v4 = main HEAD on 22 and 23 Oct; MODEL_MANIFEST.json sha256 3071f802ef11e092dfafdd72c49a616b8a288c7a0791e0966f6f2fd55021a863, 26 files, 7,976,303,844 weight bytes, 8 chart PNGs excluded and never read; 5 LFS oids and 21 git-blob oids checked, the author's SHA256SUMS agrees on all 24 non-PNG entries; all 26 files re-verified on the VM after download). Command (the card's Quick-start, all defaults): python3 /code/code/serve_head.py --snapshot /models/StandardOne-3B-SH --head /models/StandardOne-3B-SH/head --port 30171 --served-model-name standardthinking/standard-one-3b-sh (batch 8 requests / 32,768 tokens / 5 ms wait, bf16 backbone, torch SDPA, no cache). The 18 files executed or read at inference (LICENSE, SYSTEM_PROMPT.txt, chat_template.jinja, code/{head,model,render,scoring,serve_head}.py, code/requirements.txt, config.json, generation_config.json, model.safetensors.index.json, params.json, processor_config.json, special_tokens_map.json, tekken.json, tokenizer.json, tokenizer_config.json) are byte-identical (sha256) to the v2 route that was security-reviewed (PR on file / PR on file) and measured on 12 Oct 2026; only the two backbone weight shards, head/schema_head.safetensors and head/schema_head_config.json differ (the config diff is two provenance keys, step 57 -> 11471 and snapshot -> StandardOne-3B-SH@v3, which the server never reads: scoring.py:162-170). The source changes and runner were reviewed before this run; the unchanged serving code retained its earlier clearance. Image: pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime@sha256:eee11b3b3872a8c838e35ef48f08b2d5def2080902c7f666831310ca1a0ef2be plus the hash-pinned runtime supplement (19 wheels), built offline on the VM (image sha256:6d8e3c0234ba9923d2c59b85c6b046e6438a34dc1d1504a0218f9a3d00e59f12; the build is not bit-reproducible, the 12 Oct image was 28f68ce6...); transformers 5.12.1, peft 0.21.0, tokenizers 0.22.2, torch 2.11.0, Python 3.12. Container read-only, network none, model and code read-only. Weights: Apache-2.0, public, not gated, tensors in safetensors only (no pickle). FIRST MEASUREMENT OF THIS VERSION. NEW VERSION OF AN EXISTING MODEL (Author submission, free queue, filed by the authors as a follow-up to, not a replacement of, the live row standard_one_3b_sh, which measured the v2 weights at revision bead20e94b7c96045a19dc4d43af48a66c8911ea via paid order 3a081599 on 12 Oct 2026). Whether this row supersedes the live row or stands beside it is a release-owner decision. AUTHOR'S OWN SERVER, on our hardware: the live row ran on a Lium H100; this row runs on one RTX PRO 6000 (bf16 SDPA numerics can differ across GPUs by tiny amounts), so the two rows differ in hardware as well as in weights. SECOND USE OF THE DRAW: evaluator exposure and device ruling admitted this as a same-author second use of the frozen draw for the key family standard_one_3b_sh (first use: the 12 Oct v2 row), with this disclosure; any further StandardThinking row goes to the next draw. 10 s RULE: max measured decision latency 0.116 s (p50 0.031 s, p95 0.041 s, p99 0.102 s); 0 decisions over 10 s. The 15 refused items are the 15 longest (76k-82k tokens), refused with HTTP 422 `too_long` at positions 65, 169, 474, 482, 515, 573, 580, 729, 833, 899, 951, 1057, 1116, 1198, 1418 exactly as pre-registered and identical to the 12 Oct set; they are scored wrong, never truncated. DETERMINISM: first 50 positions run twice, 2/50 bit-identical, 0 argmax flips, max probability difference 9.2e-08. AUTHOR EXPOSURE AND DIRECTION (author's own release-manifest.json, self-reported, not verified): the v4 head was trained from the unsubmitted v3 (head config snapshot StandardOne-3B-SH@v3); between v2 and v4 the author reports its internal `jevbench-original` set falling 98.61 -> 95.83 and `hardproxy` 41.67 -> 38.0 while `internal-eval-v1` rises 84.33 -> 86.33 and `-v2-realistic` 88.33 -> 90.17. The author's training data and overlap screening for v3/v4 were not independently audited by us, and we ran no overlap check against the public 231. OFFLINE COUNTER: a first-party transcription (no author import) that reproduced the 12 Oct native counts exactly; it does not transcribe the span-coverage checks; the exact sealed token match with the 12 Oct receipt establishes the accounting on this draw, not full renderer equivalence. No outside request for this model, and no score, rank or expectation was communicated to the submitter. PREDECESSOR: standard_one_3b_sh (/jev-models/standard_one_3b_sh); both measurements remain separate rows.