Phocinae-Largha-150M-v1
A 144.3M-parameter local decision model released under the Apache-2.0 license, designed for approval-gate scenarios. It supports typed questions such as yes/no, choice, and rating, produces zero output tokens, and can make deterministic decisions.
Timeline (last 90 days)
The workload measured here is the public typed-decisions row set: 400 cases and 2000 decisions, with the local model being Phocinae-Largha-150M-v1 (CPU, fp32) and the hosted model used for comparison being deepseek-flash in the flash tier (temperature 0).
The decision model takes a state and up to five typed questions (noul/choice/score) and returns answers and confidence in a single forward pass, generating no text, so output_tokens is 0.
On GPU fp16 (RTX 5090), Largha-150M-v1 reports a p50 latency of 21.0 ms per decision.
On a single CPU thread, Largha-150M-v1 takes about 1.64 seconds to process one case (one state plus five questions); with 8 threads, about 8–20 decisions/second.
Largha-150M-v1 achieves a reading of 0.906 on the typed-decision English self-evaluation (400 cases / 2000 decisions), higher than the 0.735 teacher self-consistency reference, but the dataset card flags a risk of label-specific overfitting.
Largha-150M-v1 reads 0.5455 (126/231) on JevBench public-231, below its 58.4% acceptance threshold, and does not pass.
Largha-150M-v1's Chinese reading is 0.848, but the training mixture contains about 2400 machine-translated Chinese lines and about 1400 native Chinese lines, so it should be regarded as a fitting result rather than zero-shot transfer.
Phocinae-Largha-150M-v1 is a 144.3M-parameter, Apache-2.0-licensed local decision model.
Under a confidence gate of τ=0.6, Largha-150M-v1 escalated 45.0% of official-set decisions, reducing LLM calls by 55.0%, raising the retained-subset accuracy from 0.906 to 0.9936, with an overall routing accuracy of 0.8135.
The model card notes that τ=0.6 is an operating point obtained by scanning over the evaluation set, and must be re-scanned for each domain.
5 more timeline events
All Facts (15)
The workload measured here is the public typed-decisions row set: 400 cases and 2000 decisions, with the local model being Phocinae-Largha-150M-v1 (CPU, fp32) and the hosted model used for comparison being deepseek-flash in the flash tier (temperature 0).
50%UnverifiedThe decision model takes a state and up to five typed questions (noul/choice/score) and returns answers and confidence in a single forward pass, generating no text, so output_tokens is 0.
50%UnverifiedOn GPU fp16 (RTX 5090), Largha-150M-v1 reports a p50 latency of 21.0 ms per decision.
50%UnverifiedOn a single CPU thread, Largha-150M-v1 takes about 1.64 seconds to process one case (one state plus five questions); with 8 threads, about 8–20 decisions/second.
50%UnverifiedLargha-150M-v1 achieves a reading of 0.906 on the typed-decision English self-evaluation (400 cases / 2000 decisions), higher than the 0.735 teacher self-consistency reference, but the dataset card flags a risk of label-specific overfitting.
50%UnverifiedLargha-150M-v1 reads 0.5455 (126/231) on JevBench public-231, below its 58.4% acceptance threshold, and does not pass.
50%UnverifiedLargha-150M-v1's Chinese reading is 0.848, but the training mixture contains about 2400 machine-translated Chinese lines and about 1400 native Chinese lines, so it should be regarded as a fitting result rather than zero-shot transfer.
50%UnverifiedPhocinae-Largha-150M-v1 is a 144.3M-parameter, Apache-2.0-licensed local decision model.
50%UnverifiedUnder a confidence gate of τ=0.6, Largha-150M-v1 escalated 45.0% of official-set decisions, reducing LLM calls by 55.0%, raising the retained-subset accuracy from 0.906 to 0.9936, with an overall routing accuracy of 0.8135.
50%UnverifiedThe model card notes that τ=0.6 is an operating point obtained by scanning over the evaluation set, and must be re-scanned for each domain.
50%UnverifiedLargha-150M-v1 is weak at tool routing: the model card's case shows it returned fetch_url instead of web_search at about 0.33 confidence.
50%UnverifiedLargha-150M-v1 is not suitable as a safety judge, and should be used as a first-line signal together with escalation, rules and human paths.
50%UnverifiedLargha-150M-v1 has a base context of 8192 tokens, and the model card explicitly states that it is not suitable for long-document reasoning.
50%UnverifiedLargha-150M-v1's training mixture includes a pool of 500 error corrections, which were mined from the model's predictions on typed-decision test rows, so 0.906 is a fitting reading rather than a zero-shot result.
50%UnverifiedThe confidence column shipped by Largha-150M-v1 has an ECE of 0.2519, and the built-in recalibration column (temperature scaling) reduces it to 0.0168.
50%