← sources

Got jev-like api running natively on ninfer / qwen3.8 27b and results are quite decent

source: reddit (reddit:LocalLLaMA)published 2026-09-21 14:46ingested 2026-09-22 02:40

retained text

Hi all! After all the jev hype I wanted to try by myself and got qwen3.8 27b to mix-and-match open implementations on top of ninfer, and to be honest I'm quite impressed with the results, latency is not great but accuracy is almost on par, this is a very dirty POC but it kinda works (there's still some bugs and incorrect allocation, still need to figure out how to properly integrate into ninfer). Source code at: [https://github.com/knoopx/ninfer](https://github.com/knoopx/ninfer)

# Benchmark

Measured against [JevBench v1.2](https://benchmarkheaven.com/jev-models), a decision-model benchmark. The public set is 231 typed decisions (48 easy, 111 hard, 72 original): each is a `state` plus a bounded rubric, answered with a typed result and per-option probabilities. The run drives the native `POST /v1/decisions` route directly (state + question in, typed probabilities out), not the chat-completions verbalization path.

Setup: model [`https://huggingface.co/knoopx/Swift-Qwen3.8-27B-NInfer`](https://huggingface.co/knoopx/Swift-Qwen3.8-27B-NInfer) , RTX 5090, three in-flight decisions. All 231 tasks scored, 0 errors, 0 admission blocks. Wall clock 19 s.

# Headline

|Metric|Value|
|:-|:-|
|Accuracy (231 public)|84.4% (195/231)|
|ECE (10 bins)|0.045|
|Brier|0.081|
|Latency p50 / p95|127 ms / 699 ms|
|Speed axis (JevBench ×2 + 0.15 s)|\~81.7|

# Accuracy by tier and type

|Tier|choice|noul|score|total|
|:-|:-|:-|:-|:-|
|easy|100% (36/36)|100% (12/12)|—|100% (48/48)|
|original|97% (35/36)|95% (23/24)|100% (12/12)|97.2% (70/72)|
|hard|70% (47/67)|73% (28/38)|33% (2/6)|69.4% (77/111)|
|all|84.9% (118/139)|85.1% (63/74)|77.8% (14/18)|84.4% (195/231)|

# Hard tier by family

|Family|Acc|
|:-|:-|
|temporal-numeric|33% (5/15)|
|probability|50% (5/10)|
|tradeoff|67% (4/6)|
|long-policy|68% (13/19)|
|multi-hop|72% (13/18)|
|judge-hard|76% (13/17)|
|ambiguous|86% (6/7)|
|trap|88% (7/8)|
|adversarial|100% (6/6)|
|routing-hard|100% (5/5)|

# Latency (seconds)

|Scope|p50|p90|p95|p99|max|mean|
|:-|:-|:-|:-|:-|:-|:-|
|all|0.127|0.619|0.699|0.896|0.925|0.238|
|easy|0.100|0.103|0.453|0.691|0.736|0.134|
|original|0.101|0.123|0.123|0.309|0.331|0.109|
|hard|0.286|0.694|0.774|0.911|0.925|0.366|

Hard-tier p95 (774 ms) is prefill-dominated by the \~3.7k-token states. Tail latency (p95 and above) is queueing-sensitive: a single decision worker serializes execution, so a task can queue behind heavier in-flight decisions, inflating per-tier tails (visible in the easy row). p50 and mean reflect the compute cost; accuracy and calibration are unaffected by queueing.

# Calibration

36 of 231 wrong. ECE 0.045 (10 bins), Brier 0.081 — the model is well calibrated overall. The five most-confident wrong answers:

|Task|Conf|Predicted|Gold|
|:-|:-|:-|:-|
|hard-opus-b-tradeoff-01|0.949|escalate\_to\_security|escalate\_to\_engineering|
|hard-opus-a-probability-08|0.852|yes|no|
|hard-sol-b-judge-hard-02|0.852|yes|no|
|hard-opus-c-temporal-numeric-02|0.798|no|yes|
|hard-opus-a-probability-04|0.729|late|on\_time|

# Field comparison (same 231 public tasks)

Every published system's public-set accuracy, against its published JevBench composite (the composite is a geometric mean over Intelligence, Calibration, Speed, and Cost on the full 534-task set, so it is not directly comparable to a public-set accuracy). This run has no published composite (self-hosted, no Speed/Cost measurement):

|Rank|System|Score|Pub acc|
|:-|:-|:-|:-|
|1|Jev 1.13.0 (TypeSafe AI)|75.4|86.6%|
|2|SemIf, formerly OpenJev (Qwen3.5-4B)|74.7|81.0%|
|3|djev (Maisa, diffusion-gemma)|74.3|84.0%|
|4|openJev Verdict 1.4|72.5|57.6%|
|5|Laya (ModernBERT-large 421M)|70.1|58.4%|
|6|open-alternative-jev (Qwen3.5-4B)|69.8|74.0%|
|7|system-one-open (Gemma 4 E2B LoRA)|68.9|73.2%|
|\*|**NInfer ukisai/Swift-Qwen3.8-27B (this run)**|**68.4**|**84.4%**|
|8|OpenJev (DiffusionGemma 26B NVFP4)|67.7|81.8%|
|9|SimpleJev Qwen3.8-27B|67.3|86.6%|
|10|jeff (GLiFormer 400M)|66.9|62.8%|
|11|kev 0.6B|66.7|66.7%|
|12|openjev-sglang (Qwen3.6-35B-A3B)|66.3|85.3%|
|13|openJev Verdict (ModernBERT-base 151M)|66.2|55.4%|
|14|GPT-5.6 Luna (low reasoning)|66.2|97.4%|
|15|open-jev-deberta-v3-large (CPU)|64.6|52.4%|
|16|SimpleJev Qwen3.6-35B-A3B|63.8|81.4%|
|17|Bespoke Nimble 9B|63.7|67.5%|
|18|kev 0.5B|63.1|49.4%|
|19|kev 4B|62.2|66.2%|
|20|Gemini 3.1 Flash-Lite|60.9|87.0%|
|21|kev 8B|58.3|71.4%|
|22|DeepSeek V4.1 Flash (thinking)|57.8|97.8%|
|23|system-one (Qwen3-8B)|56.6|71.9%|
|24|GLiNER2 (Fastino)|53.0|58.0%|
|—|classifier.dev (fast tier)|84.8|85.3%|
|—|Qwen3.8 27B (Chutes TEE)|25.5|71.9%|
|—|Needle 3 (options as tools)|19.2|22.1%|
|—|Needle 3 (Cactus, 2-bit, CPU)|16.8|22.5%|

document fetched 2026-09-22 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wmev5u/got_jevlike_api_running_natively_on_ninfer_qwen38