TEXT01
Themis 0.8B
A compact text model for routing, triage, relevance checks and other bounded judgments.
- Input
- Text
- Rendered context
- 2,048 tokens
- Backbone
- Qwen3.5-0.8B
- Released checkpoint
- Step 4,153
Introducing Themis · Open weights · v1.0
Small models.
Clear decisions.
Turn text and images into choices, yes/no judgments and rubric scores. A probability for every option. One forward pass.
*Our evaluation; strict 4K overall: 82.54%. Read the protocol ↗
Built on Qwen3.5. Complete weights and inference code.
Apache
2.0 · Run on your own hardware.
01 / A decision in practiceRecorded · 2B
STATE
“A customer was charged twice
and wants a refund.”
THEMIS / ONE PASS
CHOICE
98.7%
Route to billing
YES / NO
84.6%
Investigation needed
SCORE
51.5%
Medium urgency
Actual release-verification output
Inspect the distributions ↗
2model sizes
3decision primitives
255options per question*
1shared forward pass
01 / In practice
Same model. Different decisions.
Read the evidence. Inspect the alternatives.
Every probability
below is a saved model output.
Loading recorded cases…
02 / The family
Pick your scope.
Two sizes. The same typed-question approach.
Choose the context
and modalities your workflow needs.
TEXT + VISION02
Themis 2B
Bring images into the decision: diagrams, document pages, screenshots and charts alongside text.
- Input
- Text + up to 2 images
- Rendered context
- 4,096 tokens
- Backbone
- Qwen3.5-2B + native vision
- Released checkpoint
- Step 3,258 · v1.0.0
*Up to 255 verified single-token answer codes, subject to context. Context includes evidence, questions, options, template and image tokens. Oversized requests raise an error.
03 / The evidence
The results, in context.
Published checkpoints, fixed evaluation corpora,
and the
details behind every number.
DECISION BENCH · 2B83.01%
889 of 1,071 cases correct
IMAGE-BEARING SUBSET · 2B93.44%
114 of 122 cases correct
READ THE PROTOCOL
Our measured results, not official leaderboard entries. Nine long cases used 8K context; strict 4K accuracy is 82.54%.
Methodology & overlap notes ↗Loading verified aggregate results…
04 / Explore the output
A decision you can inspect.
See how the probabilities move across models.
Then export a
request to run yourself.
REQUEST
PROBABILITIESThemis 2B
Precomputed during release verification. Selecting an example replays saved results; no model runs in your browser.
Build your own request +
Edit the evidence and question schema, then copy a complete Python example or download the JSON. Custom requests run on your own machine; this site has no live inference service.
05 / Your first decision
A case in.
Clarity out.
Describe your evidence and the answers you will accept. Ask several typed questions in one request, then work with the returned probabilities.
- 01 Load the published weights.
- 02 Define your case and questions.
- 03 Read one distribution per question.
Runs on your own hardware. Follow the model card for tested dependencies. The response tab shows the recorded 2B refund example.
THEMIS 2B / LOCAL INFERENCE
Loading the pinned example…
Pinned release · no hosted API required
06 / Inside the model
From evidence to probability.
A Qwen decoder adapted for typed decisions.
The answer space
is defined by your application.
-
01
Define the decision
Supply the case, questions and descriptions of the allowed options. Text and optional images enter the encoder.
state + questions + options -
02
Read the context
Full-attention layers see both directions. Linear-attention layers keep forward recurrence. Multiple questions share the request.
Qwen3.5 + decoder LoRA -
03
Score the options
Hidden states at dedicated answer positions score only valid answer-code embeddings. Softmax produces each distribution.
answer states → option scores
A small decision vocabulary.
Choice selects a named alternative. Yes/no estimates whether a statement is true. Score evaluates an ordered rubric.
Answers come with option probabilities. The 0.8B release also reports an expected score; the 2B release returns the most likely level index. Their response envelopes differ.
The image stays in the loop.
The 2B model uses Qwen’s frozen native vision tower and merger. Image tokens enter the decoder together with the case. Decoder LoRA learns the decision task; vision weights stay frozen.
Up to two images per request, with 64–1,024 visual tokens each. Those tokens count toward the 4K context limit.
Choose the checkpoint on validation.
BF16 LoRA, rank 32, with split-aware probability targets and target-preserving choice permutations. The 2B experiment completed three epochs; step 3,258 had the lowest validation soft cross-entropy.
Per-type temperatures were fitted on calibration data. Better confidence calibration cannot repair an incorrect top-ranked answer.
2B TRAINING RECIPE
- Training decisions
- 73,247
- Packed requests
- 52,111
- Precision / adapter
- BF16 · LoRA r32 / α64
- Batch / learning rate
- 32 requests · 1e−4
- Training / hardware
- 3 epochs · one H100
- Estimated experiment compute
- US$16.68*
*Recorded estimate including preparation, profiling and evaluation; not a guaranteed invoice or a serving-cost estimate. Base weights and embeddings stayed frozen; decoder LoRA was trained with soft cross-entropy.
07 / What the numbers mean
Keep the boundaries visible.
Results are useful when the protocol travels
with them. Here
are the important qualifications.
Context & coverage
Decision Bench has 949 text and 122 image-bearing cases. The headline 2B run extended nine long requests to 8K. Counting those cases as failures yields 82.54%; the released model enforces 4K by default.
The 0.8B text comparison used an 8K limit. Its strict trained 2K result is 70.60%, compared with 72.50% when extended.
Overlap & independence
Typed Decisions is a fitted evaluation: 554 official training cases enter the 2B mixture. No official test states matched train, validation or calibration by exact normalized state checks.
172 official test cases also appear in the mixture test split. Those evaluations are not independent. Paraphrase overlap and base-model pretraining exposure remain unassessed.
Probabilities & performance
Brier score and ECE definitions vary across benchmarks. Compare checkpoints within a protocol. SysOne scores expected rubric levels, while Typed Decisions uses the highest-probability answer.
Local latency is workload- and hardware-specific. It is not hosted API latency, and there is no controlled speed comparison against Jev or other vendors.
Where judgment still matters
Themis scores the alternatives you provide. A missing correct option cannot be selected. Question bundling and option order can affect the result.
English prompts were used for training and evaluation. New domains, other languages, quantization and longer contexts need independent validation. High-stakes outcomes require review.
Why publish step 3,258 instead of the final checkpoint? +
08 / Put it to work
Your questions.
Your decision space.
Start with complete weights and a runnable example.
The full
curation, training and evaluation pipeline
is being polished
for a future source release.