Jev directory
CommunityJev Models & Reimplementations
Torchcast Decision 4B
Torchcast AI's 4B text decision model on Qwen3.5-4B returns yes/no probabilities, choice distributions and ordinal score distributions through a TypeSafe-compatible API, reading option letters in one forward pass under CC BY-NC 4.0.
- Reading time
- 2 min
- Published by
- Community
- Author
- Torchcast AI
- Last verified
- Oct 2026
- Capabilities
- Self-hostedRuns locally
Highlights
- Decision model on Qwen/Qwen3.5-4B returning noul, choice and score distributions without generating text.
- Reads option letters at the answer position in one forward pass; choice lists over 26 options take a few passes over groups.
- Serving config pins calibration temperature 0.9 and a yes/no commit floor of 0.801; prompts over 16,384 tokens keep head and tail.
- Pinned Triton kernels make repeated runs on the same GPU produce identical probabilities from a cold or warm cache.
- Weights are CC BY-NC 4.0; the bundled serving package is Apache-2.0 derived from StartLux-Decision inference code.
Watch out
Card reviewed 2026-10-17; the card publishes no benchmark numbers of its own and warns probabilities can be unreliable on new domains. The weights are CC BY-NC 4.0, so commercial use is restricted, and the card stresses that generic Hub text-generation examples do not reproduce the evaluated interface, only the bundled serving package does.
Questions about this entry
Short answers drawn from the linked project or page.
A text-based decision model by Torchcast AI built on Qwen3.5-4B. It returns yes/no probabilities (noul), choice distributions and ordinal score distributions through a TypeSafe-compatible API, with nothing generated.
Each decision reads the option letters at the answer position in one forward pass, and choice lists with more than 26 options are decided in a few passes over groups; the request and response format is TypeSafe-compatible.
With the bundled torchcast_decision serving package and its decision_config.json; the card warns that generic Hub text-generation examples do not reproduce the evaluated interface. The pinned artifact is tag v1.1.0, served as torchcast-decision-4b-v1.1.
Calibration temperature 0.9 and a yes/no commit floor of 0.801, below which the maximum of P and 1 minus P is moved to 0.801 on its own side without changing the answer; prompts over 16,384 tokens keep the head and tail.
None. The card carries no accuracy, calibration or latency numbers, and it states that probabilities can be unreliable on new domains.
Weights are CC BY-NC 4.0 (non-commercial), the Qwen3.5-4B base stays Apache-2.0, and the serving package is Apache-2.0 derived from StartLux-Decision inference code; the project is independent of TypeSafe AI.
English text-based decision research such as classification, routing and bounded rubric evaluation. The server returns decisions and does not execute actions, and the card notes the project is independent of the JevBench maintainers.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
chenz53's standalone decision model derived from Google Gemma 4 26B A4B IT: full merged BF16 weights, a fitted temperature of 1.736507, and 79.50% equal-type accuracy over 1,100 final questions.
Jev Models & Reimplementations#community#open-models#local
Community
A pointer-network decision model from Keyush Nisar that reads one agent state document plus typed questions and returns calibrated tool-routing probabilities in a single forward pass, reporting 89.00% zero-shot tool routing.
Jev Models & Reimplementations#community#open-models#local
Community
StartLux's 9B Instant decision model as GGUF for llama.cpp: Q8_0 (9.53 GB) measures DecisionPivot 53.2 against 50.6 for Q4_K_M and 53.2 for BF16 on vLLM, returning one-forward-pass typed-question probabilities.
Jev Models & Reimplementations#community#open-models#quantized
Community
A Strands decider v1 checkpoint: a LoRA adapter and readout head on google/gemma-4-E4B-it that answers typed decisions with calibrated confidence; its model-index reports 0.8095 JevBench public accuracy (187/231) and mixed code-bug evaluation scores.
Jev Models & Reimplementations#community#open-models#local
Community
A Strands decider v1 checkpoint: a LoRA adapter and readout head on google/gemma-4-E2B-it that answers typed decisions with calibrated confidence; its model-index reports 0.7792 JevBench public accuracy (180/231) and code-bug evals from 0.281 to 0.978.
Jev Models & Reimplementations#community#open-models#local
Community
Jared Palmer's 27B Kev decision model: a full-weight fine-tune of Qwen3.8-27B averaged with v1, plus a pointer head that returns calibrated probabilities for typed choice, score and noul questions in one forward pass. Card-reported transfer-v4 accuracy is 88.87%.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.