01
TOD on τ-bench
Took over 77% of actions while holding the frontier model’s solve rate, taking the same action as models 100x its size.
Solutions · Decision models
If It’s Worth Deciding,
It’s Worth Fine-Tuning.
Decision models read like an LLM and answer like a classifier, in one forward pass. Parsec tunes them on your decisions and serves them at scale.
- per 1,000 decisions
- $0.30
- per 1,000 decisions
- typical tuned improvement
- 15%+
- typical tuned improvement
- to a post-trained model in production
- 2 weeks
- to a post-trained model in production
From old to new
A classifier that reads every token.
Decision models are a special kind of classifier. Given input X, pick 0 or 1, a score, or one of many labels.
They read every token with full attention, like an LLM, then put their weight on one of the given options. No generation, so one batched forward pass. Much cheaper and faster.
Why fine-tune
Fast and cheap means inflexible.
The market-leading decision models are non-reasoning. They lean on their trained priors, and those priors are someone else’s assumptions.
Better context helps, but it is not a model you own. Decision models are among the easiest to tune: simple hardware, often a few thousand examples. Gains are often double digits, and on novel tasks they are the difference between having the capability or not.
Our open-source fine-tuning SDKWhat we have built
A few recent projects.
02
Ragzilla
A small ModernBERT that judges retrieved chunks before the LLM sees them. 98% of evidence kept, 80% fewer tokens.
03
TOD2 on 10-Ks
Long documents where the relevant facts are spread out. Double Jev’s performance on our held-out set.
04
TOD plays Papers, Please
No internal game state, just the screen, a mouse and a keyboard. It beat the game.
Where Parsec fits
The evals, the gyms, and the GPUs.
Parsec serves your traffic, then generates the evals, training sets and post-training environments to improve your decision models. TOD is built to be tuned.
A 10% better decision can drive 10–100x the value. One misclassified refund on a delivery order costs about $40. Make it 1,000.
Decision models are too big for a CPU and too small to keep a GPU busy. Self-hosting one means paying for an idle card. We serve them at scale.
The simple truth
Parsec vs other vendors vs self fine-tuning.
| Parsec | Other vendors | Self fine-tune | |
|---|---|---|---|
| Cost | $0.30 per 1,000 decisions | $0.40 per million tokens | Always-on GPUs |
| Accuracy | Tuned by experts, typically 15%+ improvements | Whatever they give you | The best you can manage |
| Post-training | Real gyms | Not a chance | 20–40 hours to stand up |
| Maintenance | Always improving, no effort on your end | None, and it shows | Constant drag, spiky updates |
| First-time effort | One-line switch | Zero | 100–200 hours for an experienced engineer |
Point Your Inference. We Do the Rest.
A post-trained decision model, tested in a gym built for your business, in production within two weeks.