Torchcast Decision 4B (v1.3.1)

A text-based decision model developed by Torchcast AI, based on Qwen3.5-4B. It returns yes/no probabilities (noul), choice distributions (choice) and ordinal score distributions (score) through a TypeSafe-compatible API. Each decision reads the option letters at the answer position in one forward pass (choice lists with more than 26 options are decided in a few passes over groups of options); nothing is generated.

Use the bundled torchcast_decision serving package with the bundled decision_config.json. Generic Hub text-generation examples and causal-attention runtimes (vLLM, llama.cpp, MLX) do not reproduce the evaluated interface: in this version the full-attention layers attend to the whole prompt in both directions (see below).

Serving instructions

What changed in v1.3

• Weights: a new full fine-tune of Qwen/Qwen3.5-4B on an expanded training mix of public datasets (official training splits) and synthetic decision data covering choice, yes/no and score questions. The architecture, tokenizer, prompt format (jev_render_v3) and attention mode are the same as in v1.2.
• Speed: the bundled server now uses CUDA graphs; median latency per request is about half that of v1.2 (67.9 ms to 31.9 ms on one A40), with the same probabilities as the eager path.
• Memory: about 1.2 GiB less GPU memory after loading.
• Serving settings: updated in decision_config.json.

v1.3.1 (serving package only): the weights and decision_config.json are those of v1.3.0, and the probabilities are identical (bit for bit) to v1.3.0's. The serving package torchcast_decision/ was rewritten by Torchcast AI; the server now answers every error with JSON and a status code (503 when the GPU runs out of memory, 500 for anything else) instead of closing the connection. The Apple Silicon (MLX) and GGUF backends, which could not run this model, were removed.

Results

On the JevBench leaderboard (release v1.7.9, as of 2026-10-27), Torchcast Decision 4B v1.3 is #1 among open-weight models by composite score:

Source: benchmarkheaven.com/jev-models. Measured by the JevBench maintainers at revision 004acea2 (tag v1.3.0; v1.3.1 has the same weights and probabilities) with the bundled server and settings, on the second regular v1.6 draw (anchor runs pending, so scores may be re-placed when the board's anchor offsets are applied).

Artifact

• Pinned revision: tag v1.3.1 of this repository (the board measured tag v1.3.0, same weights).
• Weight SHA-256: see SHA256SUMS.
• Served model name: torchcast-decision-4b-v1.3.

Serving settings

Serve the model through the bundled torchcast_decision package; decision_config.json holds the serving settings. Within a prompt, the full-attention layers use no causal mask (as in training), so loading the weights with plain transformers, vLLM, llama.cpp or MLX gives different probabilities. Prompts longer than 16,384 tokens keep the head and the tail of the state. Repeated runs on the same GPU type give identical probabilities. See NOTICE and the serving instructions.

Intended use and limitations

English text-based decision research, such as classification, routing and bounded rubric evaluation. The server returns decisions and does not execute actions. Probabilities can be unreliable on new domains.

Licence and attribution

Model weights: Apache License 2.0 (see LICENSE). Base model: Qwen3.5-4B by the Qwen team, Apache License 2.0, Copyright 2026 Alibaba Cloud (LICENSE-qwen-Apache-2.0, NOTICE). Serving package: Apache License 2.0, written by Torchcast AI (torchcast_decision/LICENSE, torchcast_decision/NOTICE). This project is independent of TypeSafe AI and the JevBench maintainers. Earlier releases keep their own licence: v1.1 (tag v1.1.0) remains under CC BY-NC 4.0.