Phocinae-Largha-150M-v1
Data: JevBench v1.6.39, published 2026-10-22
Capability 30.2 · rank #129 of 139 on the open-weights board
Last measured: 2026-10-13 · measurement release: v1.6.15
Apache-2.0 (LICENSE + NOTICE in the weights repo; the author's phocinae-server is Apache-2.0 too)
intelligence
3.6
calibration
56.9
speed
77.1
cost
95.3
Composite (secondary)
0.1 · Composite rank #180
Cost and measurement conditions
$0.0014 per 1,000 decisions (estimated)
ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff. A 144.3M ModernBERT encoder with no decoder lineage, so the base-size-encoder line is used unchanged - USD 0.005 per 1M input and 0.0 per 1M output, the line the PUBLISHED v1.6 field already carries on modernbert-jev-v2, tacet-sonata and watt-flash-0.1 (verified in their own registry entries; modernbert-jev-v2 was measured in v1.6.8 on the earlier v1.6-regular-20261010-a2 draw, while tacet-sonata and watt-flash-0.1 are v1.6.1 rows - none of the three was measured on THIS draw, which carries no ModernBERT row), backed by the dated 25 Sep 2026 DeepInfra snapshot sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f in which thenlper/gte-base, BAAI/bge-base-en-v1.5, intfloat/e5-base-v2 and sentence-transformers/all-mpnet-base-v2 all price at USD 0.005 per 1M input and 0 output - and no new release-lane pricing line is created. Output tokens are a literal 0 reported by the author's own server (an encoder read-out; nothing is generated). x this row's own measured input tokens, which are the author's server's own count. Estimated, not charged. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.001434 per 1,000 decisions (0.022x the official cost cap USD 0.064595), against USD 0.001440 over all 1,500 items (v1.6.29).
Measured p50 latency: 0.443 s (on-CPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 1.04 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.443 s on-CPU, no network (timed on our CPU host over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 1.04 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Sandy, the evaluator's own shared CPU host (AMD Ryzen 5 3600, 6 cores / 12 threads, Zen 2, AVX2), NOT a dedicated machine: other jobs ran beside it and the 1-minute load average was logged every 60 s into logs/CONDITIONS-<key>.jsonl - for this row's window 3.75 to 9.89, mean 5.81 over 12 samples, the quietest of the three new rows. The author's own phocinae-server 0.1.5 from PyPI, UNMODIFIED, run with their own documented line `PHOC_MODEL_DIR=<pinned snapshot> python -m phocinae.main` at PHOC_DEVICE=cpu, PHOC_THREADS=1 (their own documented CPU setting: 1.64 s per 5-decision case single-threaded), PHOC_EXTENSIONS=1 for their native option_scores, PHOC_PERM_AVG at its default 0 and no bearer token, inside a loopback-only network namespace with an outbound probe that failed as required and no credentials in the environment. Their package omits numpy from its declared dependencies although their own engine.py calls .numpy(); numpy 2.5.4 was installed so their code could run, and no file of theirs was edited. OUR shim (r107/src/phoc_shim_r107.py, byte-identical to the already source-reviewed r92/src/phoc_shim_r92.py, sha256 631cb60dde86b8e609382a32900db5e4ddc9fa9b2fef39bdb7b022dff558775e) is the only translation layer in front of it, because their published API carries no question text - see the mapping disclosure. Speed is a loaded-shared-CPU figure; the standard self-hosted x2 + 0.15 s adjustment is applied on top and the raw p50/p95 are kept. Re-measured for this release on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b) under CoS deputy decision #18209, at USD 0 on Sandy’s own CPU, 1 thread, 2026-10-13 05:59:36Z–06:11:07Z, 1-minute load average 3.75–9.89 (mean 5.81, 12 samples). DISCLOSED MAPPING: choice options are sent as "<label>: <description>" in the item’s own criteria order and the author’s option_scores are mapped back onto the bare labels by position; for noul only the author’s p(true) is forwarded and the stock adapter derives p(no) = 1 - p(yes), so their own rounded p(false) is unused. The author’s API answers score on a fixed 9-level scale 2..10 against items with 4 or 5 levels and forbids score options, so no identity mapping exists and this row uses pre-committed unsupported-native-score handling fixed before any measurement; all 375 score items of this draw are recorded as item-level misses (HTTP 422), and support.score is left at "native" rather than "unsupported" so that the type is still weighted and the misses count — the stricter of the two treatments, and the same rule every other row gets.