JEV-27B
Open-weight decision model built on a frozen Qwen3.8-27B backbone. It returns calibrated probabilities for yes/no, multiple-choice and 0-5 score decisions in a single forward pass, with ordinary text generation also available.
From the model card
Specifications
Published by the model author. Reviewed 2026-10-11.
Publisher model card- Architecture
- Frozen Qwen3.8-27B backbone + System-1 decision block
- Backbone parameters
- 25.6B
- Trained decision parameters
- 108.9M
- Weight precision
- BF16
- Published context
- 262,144 tokens
- Decision options
- 2-256
- Input → output
- Text → calibrated decision or text
- License
- Apache 2.0
Measured by EcoHash
Performance on one RTX PRO 6000
Recorded under the conditions below. Single-request speed and aggregate throughput describe different workloads, and concurrency is the level that was tested, not a capacity limit.
Speed
- End-to-end latency P50
- 1.12 s
- End-to-end latency P95
- 1.25 s
Reliability
- Success rate
- 100%
- Failures
- 0
- Out-of-memory failures
- 0
Source & use
License & links
Where this model comes from and what you may do with it.
Commercial use is permitted under Apache 2.0, subject to its notice and other requirements.
- Publisher
- AutoTrust AI
- Model ID
- jev-27b
- License
- Apache 2.0
- Model card / source
- AutoTrust AI on Hugging Face
Specifications reviewed 2026-10-11. Measurements are EcoHash’s own, recorded with their full methodology below.
Measured by EcoHash
Methodology & provenance
How each number on this page was produced, with the raw archives that back it.
Recorded test configuration & full methodologyrevision 1 · updated 2026-10-27
- Hardware SKU
- EcoLink ATL model instance; RTX PRO 6000 class GPU
- Deployment region
- ATL
- Test date
- 2026-10-26
- Warmup runs
- 2
- Sample count
- 20
- Input tokens
- 46 prompt tokens in recorded response fixture
- Output tokens
- 1 decision token
- Test context limit (tokens)
- Short decision prompt; full 256K context not exercised
- Tested concurrency levels
- 1 and 20
- Workflow revision
- capacity_test.py decide category; jev-27b-decide alias
- Raw results URL (HTTPS)
- https://github.com/ecohash-ai/ecohash-benchmarks
- Test notes
- Single-pass /v1/decide benchmark. At concurrency 20, 20/20 requests passed: latency p50 1.12 s, p95 1.25 s, 22.33 requests/s (1,339.8 completed decisions/min), zero errors. The Throughput field records completed decisions/min. Chat C=1 was also exercised but TTFT p95 2.815 s exceeded the interactive gate; it is not combined with this decision result.
Updated 2026-10-27 · revision 1