Torchcast Decision 12B
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.3/10
Open weights but CC BY-NC 4.0 (non-commercial). Detailed card, pinned v1.0.0 tag, weight hash, serving repo with run manifests, /v1/systemone over vLLM 0.30.0. Community GGUF. Small company, no SLA.
- Capability8.1/10
Official JevBench (17 Oct): Intelligence 60.5, calibration 82.9, p50 0.037 s. Noul / Choice / Score by option-letter logprob readout.
- Adoption3/100
~370 downloads / 2 likes on the weights; a community GGUF has ~8.6k downloads. No launch post, integrations or production users found.
Vendor claims
- JevBench v1.4 public set: 203/231 (87.88%), author-run, not an official score; public results influenced checkpoint and serving selection[Vendor claim — not independently verified]
- Serial latency 30 ms p50 / 43 ms p95 on one H100 (easy + original items); not the official Speed measurement[Vendor claim — not independently verified]
publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.
Read this first
- Non-commercial. The weights are labelled CC BY-NC 4.0, and the card says the label “does not establish clearance of every training-source right”. Gemma’s prohibited-use policy also applies. The serving code is MIT.
- The 87.88% is not a held-out number. Public JevBench results were used to pick the checkpoint and serving settings. The independent number is Benchmark Heaven’s.
- Built on Cygnet and NInfer. The serving and readout foundation comes from Cygnet and NInfer; the weights are a real LoRA fine-tune, unlike Cygnet’s frozen base.
What it is
A LoRA fine-tune of Gemma-4-12B-it by Torchcast AI, trained on a 50/50 mix of gold labels and a larger instruction model’s option distributions (not Jev outputs). Each decision is one forward pass that reads the option letters at the answer position. Per-type temperatures sit in readout_config.json (Choice and Score 1.0, Noul 0.2).
Specs
| Attribute | Value |
|---|---|
| Weights | torchcast-ai/torchcast-decision-12b, tag v1.0.0, 12.0B params |
| License | CC BY-NC 4.0 (weights); MIT (serving code) |
| Published | |
| Serving | POST /v1/systemone, vLLM 0.30.0, BF16, one H100 80 GB (also L40S 48 GB) |
| Context | 16,384 tokens including the rendered request (HTTP 422 above) |
| Language | English |
| Sibling | torchcast-decision-27b (CC BY-NC, base StartLux-Decision-27B): Decision Index 0.3 rank 2 group (59.91), JevBench capability 81.6 but outside Jev-class (6.5× cost) |
JevBench v1.6.1 (verified 7 Oct 2026)
Open-weights board: composite 69.9 (#2), 95% interval 67.2–71.5. Intelligence 60.5, Calibration 82.9, Speed 91.7, Cost 56.4. Capability 71.7 (#5). $0.028 per 1,000 decisions (0.88× Jev), raw p50 0.22 s. Not listed on Decision Index 0.3.
Fit / anti-fit
Fit when you need: a 12B open decision model for research or internal non-commercial work, on one GPU, with a reproducible serving recipe.
Anti-fit when you need: commercial use, non-English input, more than 16K tokens, or strong calibration on yes/no questions.
Why it is in the catalog
Decision-trained weights with a typed /v1/systemone interface, documented readout and temperatures, and the #2 spot on an independent board. The license limits it to non-commercial use, and that shows in maturity.
Score working (7 Oct 2026)
- Maturity = availability 6 × 0.30 + docs 8 × 0.25 + integrations 5 × 0.25 + support 1 × 0.20 = 1.80 + 2.00 + 1.25 + 0.20 = 5.25 → 5.3
- Capability = types 10 × 0.30 + calibration 6 × 0.25 + latency 8 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.50 + 2.00 + 1.60 = 8.10 → 8.1
- Adoption = engagement 3 × 0.35 + ecosystem 5 × 0.35 + production 0 × 0.30 = 1.05 + 1.75 + 0.00 = 2.80 → 3
Availability is 6 instead of 8 because of the non-commercial license. Calibration is 6: the method is documented, but the board’s calibration axis (82.9) is among the lowest of the top 10.
Torchcast Decision 4B v1.1 (12 Oct 2026)
torchcast-decision-4b is a full-weight Qwen3.5-4B decision model (served as torchcast-decision-4b-v1.1) with a TypeSafe-compatible API and the same option-letter readout; choice lists over 26 options take a few passes. Prompts over 16,384 tokens keep the head and tail of the state. Weights CC BY-NC 4.0 (non-commercial); the serving package is Apache-2.0 and the Qwen base keeps its Apache-2.0 notice. The card publishes no accuracy numbers, and it is not on JevBench v1.6.x or Decision Index 0.3 (the board lists Torchcast Decision 27B at 59.91). Scores on this page are unchanged.
Board audit (17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.
| Board | Row | Result |
|---|---|---|
| JevBench open weights (live 17 Oct) | torchcast-decision-12b | Capability 71.7; Intelligence 60.5; calibration 82.9; p50 0.037 s; $0.028/1k |
Version / family note. This page covers the 12B only. The board also lists Torchcast Decision 27B (Decision Index Full 59.91, #19) and, on JevBench, torchcast-decision-27b and Torchcast Decision 4B; a family page is a candidate.
Score update (audit, 17 Oct 2026). Capability 8.1 → 8.1 (unchanged) = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 6 × 0.25 + 8 × 0.25 + 8 × 0.20 = 8.10 → 8.1. Types, calibration and latency come from the earlier working (10 / 6 / 8); only accuracy is re-derived from the board (8; was 8): JevBench Intelligence 60.5 ÷ 63.6 × 8 = 7.6, half-up. Maturity and adoption unchanged.
Hugging Face on 17 Oct (torchcast-ai/torchcast-decision-12b): 639 downloads, 3 likes, licence tag cc-by-nc-4.0, last modified 2026-10-02.
Torchcast Decision 4B v1.2 on JevBench (19 Oct 2026)
publicado (OK do Fabricio 19/10 BRT).
The open-weights board (draw v1.6.29) added a row for Torchcast Decision 4B v1.2, and it ranks #2 on the official ranking, behind H2O-Lightning-4B v1.1. The v1.1 row is still listed.
| Row (draw v1.6.29) | Capability | Intelligence | Calibration | p50 | Cost / 1,000 decisions | Official rank |
|---|---|---|---|---|---|---|
| Torchcast Decision 4B v1.2 | 72.6 | 60.3 | 85.0 | 0.062 s | $0.020 | #2 |
| Torchcast Decision 4B v1.1 | 68.7 | 54.2 | 83.1 | 0.021 s | $0.020 | #9 |
| torchcast-decision-12b (this page) | 71.7 | 60.5 | 82.9 | 0.037 s | $0.028 | #8 |
The 4B v1.2 now matches the 12B on Intelligence (60.3 vs 60.5) with better calibration, at a lower cost per decision. This page still covers only the 12B, so its scores are unchanged; the 4B result strengthens the case for a family page.