THX-01 322M — VRAM & /v1/systemone setup
Written by Jakub Rusinowski · Last updated
HAL-X AI's 322M multilingual decision model (mmBERT-base encoder plus a decision head). Adds number, excerpt and citation questions to choice, yes/no and score; about 10 ms a request on one GPU (author's figure).
THX-01 322M needs about 1 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
Call THX-01 322M
Served by thx01.server (default port 8095). This model does not run in Ollama.
Provisional — no independent benchmark yet. Every score on this page comes from the authors.
- Question types
- choice noul (yes / no) score number excerpt citation
- Inputs
- text
- Serve with
thx01.server- Endpoint
/v1/systemone- Languages
- Post-trained in 18 languages (Azerbaijani emphasis); 100+ supported by the encoder
pip install "thx01[server]"THX01_MODEL=doofz/THX-01 THX01_API_KEY=your-key python -m thx01.server --port 8095http://localhost:8095/v1/systemone
Setup for servers other than Ollama →
Build a request for this model Decision models guide
The builder writes choice, noul and score questions only; THX-01's number, excerpt and citation questions are not generated here. The port is the one in the model card's example.
Related guides: Confidence, thresholds and calibration · Servers other than Ollama
Other decision models: Nimble 9B · Tev1 4B · Tev1 0.8B · Winnow 12B · Winnow E4B · Decider 2B · Decider 4B · Decider 35B-A3B (NVFP4) · JevK5 4B · Intern-Decision 4B · AutoJev 27B · Laya · Clef 27B · Clef-Flash 9B · Rune 26B-A4B v3 · OpenDecider-nano · OpenDecider-small · OpenDecider-small MLX 4-bit · Bosun v3.1 0.6B · Bosun v3.1 1.7B · d1-3B · d1-omni-600M
Hardware fit
Weights plus overhead plus the KV cache at a 1,024-token prompt, on NVIDIA RTX 4090 (24 GB). A publisher build is sized from its file; the other rows are modelled at a standard quant. Decision requests are short, so no long-context figure is shown.
| Quant | Memory | VRAM | Fit |
|---|---|---|---|
| F16 Publisher build · 0.64 GB file | 1.5 GB | Fits | |
| Q4_K_M 4.83 bpw · modelled quant | 1 GB | Fits | |
| Q6_K 6.56 bpw · modelled quant | 1.1 GB | Fits | |
| Q8_0 8.50 bpw · modelled quant | 1.2 GB | Fits |
Published file size: F16 0.64 GB. A download size from the model publisher — not a VRAM requirement.
Input is capped at 1,024 tokens per question; a longer state is truncated. HAL-X also runs a free, rate-limited hosted API (api.hal-x.ai/docs/thx-01), per its launch post.
How it was scored
Each table is one suite. Numbers in different tables were measured on different questions and are not comparable, with each other or with the Decision Index.
Decision Index 0.2.1 (snapshot 2026-09-28)
Decision Index 0.2.1Community Decision Index (not by the model's authors)THX-01 is not on the Decision Index 0.2.1 snapshot (28 September 2026), and its authors have not published an Index score. Community-maintained; not affiliated with the model authors.
Source ↗| What was scored | Score | ECE | Source |
|---|---|---|---|
| 2,843 tickets in Azerbaijani, Russian, English and Turkish — average of 4 test sets | 98.4% | 0.003 | Source ↗ |
ECE is measured on the Independent test set only.
| What was scored | Score | Source |
|---|---|---|
| Number lookup (exact value) | 93.4% | Source ↗ |
| Excerpt extraction | 84.1 F1 | Source ↗ |
| Supporting citation (top part) | 94% | Source ↗ |
| What was scored | Score | ECE | Source |
|---|---|---|---|
| Mean over 39 held-out tasks | 84.2% | 0.066 | Source ↗ |
Published latency
| Device | Time | Condition | Who measured it | Source |
|---|---|---|---|---|
| One GPU (model not stated) | 10 ms | 1 request, local GPU. Jev 1.13 measured 331 ms through its API on the same set: different serving conditions | Authors' own internal benchmark — not independently reproduced | Source ↗ |
| One GPU (model not stated) | 1 ms | Batched, time per decision | Authors' own internal benchmark — not independently reproduced | Source ↗ |
As published, on the device and under the condition shown. A decision model answers in one forward pass, so this is time per decision, not tokens per second.
Decision models are not ranked on chat, creative or coding scores.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source. Still unconfirmed: hostedApi.
- Parameters
- 322M
- Context window
- 1,024 tokens per question
- Architecture
- Fine-tune of jhu-clsp/mmBERT-base
- Provider
- HAL-X AI
- Licence
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Edge device (phone, SBC)
- Specified at
- Q4_K_M
- System RAM
- 8 GB
- Record updated
- 2026-10-13
Commercial use permitted. No usage restrictions beyond attribution.
THX-01 322M — frequently asked questions
What is THX-01 322M?
THX-01 322M is a decision model: you send it a state and typed questions (choice, yes/no/unknown, or a score) and it returns one answer per question with a probability for every option. It is not a chat model.
How do I run THX-01 322M locally?
Served by thx01.server (default port 8095). This model does not run in Ollama. Start it with: `pip install "thx01[server]" && THX01_MODEL=doofz/THX-01 THX01_API_KEY=your-key python -m thx01.server --port 8095`.
How much memory does THX-01 322M need?
About 1 GB for the weights plus overhead at Q4_K_M, before the prompt's KV cache. Decision prompts are short, so the cache stays small.
How accurate is THX-01 322M?
It has no community Decision Index score. Every score on this page comes from its authors, on the suites named beside each table; none is an independent result.