# Is THX-02 SOTA?

Pending benchmarks. Announced Nov 5, 2026; no independent results yet.

As of Nov 22, 2026, THX-02 is pending benchmarks on isitsota: announced Nov 5, 2026, with no independent results on the boards we track yet.

HAL-X's open-weight multimodal decision model: a 4.5B decision network plus a 60M think adapter that returns calibrated probabilities over options you supply (choice, noul yes/no, and score). Accepts text and up to four images; trained on requests up to 32K tokens. Successor of THX-01 (0.3B). Apache-2.0 weights on Hugging Face (HAL-X-AI/THX-02) and ModelScope (HAL-X/THX-02). Hosted API at api.hal-x.ai (free-tier per IP and keyed); pip package thx02 for the API client and local/Termux runs.

Canonical: https://isitsota.com/m/thx-02
Last updated: 2026-11-05
Tracking since: 2026-11-05

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | HAL-X |
| Announced | Nov 5, 2026 |
| API | Live · Hosted at api.hal-x.ai (model id thx-02); free tier without a key (THX-02: 60 decisions/min per IP) and keyed access; docs at api.hal-x.ai/docs/thx-02/ |
| Open weights | Published |
| Parameters | 4.5B decision model + 60M think adapter |
| Role | Non-autoregressive decision model: answers typed questions over a state with a probability for every supplied option; free-text only when the think adapter writes a short note |
| Question types | choice (up to 255 options), noul (yes/no), score (ordered scale); up to 64 questions per request |
| Context | Trained up to 32K tokens per request |
| Modalities | Text plus up to 4 images per request; post-training and evaluation across more than 20 languages |
| Thinking | Optional 60M think adapter with self-gate (think=false, auto, or true); auto mixes direct and thought-assisted reads |
| Checkpoints | HAL-X-AI/THX-02 on Hugging Face and HAL-X/THX-02 on ModelScope (BF16; FP32 on CPU); includes thinker/ adapter for vLLM |
| Local | pip install thx02 (API client/CLI) or thx02[local]; Termux path documented on the model card |
| License | Apache-2.0 |
| Predecessor | THX-01 (0.3B; doofz/THX-01 on Hugging Face) — not a tracked page on isitsota |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: on a 6,018-request sample of the Decision Index 0.3 public suite, THX-02 scores 65.06 with think=auto and 63.83 direct (lab-measured on one RTX PRO 6000).
- Lab claim: 72.19 on the Decision Vision Index 0.3 public sets run so far (10 of 11; Winoground gated; no thinking).
- Lab claim: median latency 50 ms direct and 79 ms with self-gated thinking on one RTX PRO 6000; 129 requests per second on one RTX 4090 (lab-measured).

## Sources

- REPO: [HAL-X-AI/THX-02 model card (Hugging Face)](https://huggingface.co/HAL-X-AI/THX-02)
- REPO: [HAL-X/THX-02 on ModelScope](https://modelscope.ai/models/HAL-X/THX-02)
- LAB: [THX-02 API reference (HAL-X)](https://api.hal-x.ai/docs/thx-02/) (2026-11-05)
- LAB: [HAL-X announcement on X](https://x.com/halx_ai/status/2118377213842379072) (2026-11-05)
- PRESS: [ModelScope announcement on X (host)](https://x.com/ModelScope2022/status/2118372777107599433) (2026-11-05)

Tracking continues until an independent board lists it.

## Related

- Lab: [HAL-X](https://isitsota.com/labs/hal-x.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/thx-02
