Kaer-R1 7B is the first release in the Kaer Reasoners line. It is a dense 7B model with a learned halting head that decides, per question, how much deliberation the question deserves. Full method, results and failure analysis are in R-09; these are the operational notes.
What shipped
- Weights — bf16 and 8-bit builds, staged access.
- Halting head — the classifier that scores, at nine checkpoints through a reasoning trace, whether continuing will change the answer.
- Training recipe — budget-matched supervision over 148,000 problems sampled at five thinking budgets.
- Evaluation card — including all six halting-head variants that performed worse than no halting head at all, and the paraphrase degradation below.
Numbers
| Metric | Value | Conditions |
|---|---|---|
| Hard-reasoning suite | 61.4% | Single greedy pass, 5 seeds, mean |
| Mean thinking tokens | 790 | Adaptive; 4,096 if budget fixed |
| Cost / 1k items | £3.10 | Measured, not estimated |
| Abstention AUROC | 0.83 | In-distribution |
| Abstention AUROC | 0.68 | Paraphrased items (R-07 families) |
| Error reduction via routing | 41% | Bottom decile escalated |
Known limitations, in full
- Paraphrase fragility in the halting head. Reword a question and its sense of its own difficulty degrades faster than its accuracy does. It under-thinks confidently. Unfixed.
- Easy-but-wrong. On the 17% of items it gets wrong at every budget, abstention fires weakest exactly where it should fire hardest.
- Latency variance. Per-item budgeting means per-item variance. Services with a hard latency ceiling should set a fixed budget and take the lower score.
- Single-turn only. Multi-turn budget allocation is the next study; there are no results yet.
- Quantisation changes behaviour. Our 8-bit build scored within noise on capability and measurably worse on hedging and refusal manner. Re-evaluate the build you actually serve.
- No safety hardening. Beyond the base model, none. Not for public-facing deployment.
What it cost us
11,400 A100-hours across trace generation, supervised fine-tuning and eight halting-head variants — about £19,000. Pre-training used an existing open-weight base; we did not train from scratch. The single largest line was generating budget-response curves at five budgets, 62% of the bill, which in hindsight could have been three budgets for most of the corpus.
Getting it
Staged access via [email protected]. The hard-reasoning suite stays private for the usual reason: publish it and it is in the next crawl inside a month, after which the numbers mean nothing to anybody. Protocol, rubric and per-model results are open.