Primary use stated or inferred cautiously from official documentation.
Ai2
AI2 / allenai/Bolmo-1B-Stage1
Bolmo 1B Stage 1
Ai2 checkpoint included in the public release of October 7, 2026, alongside the publication of Bolmo in Nature. It is an autoregressive model that processes text as UTF-8 bytes through local components, dynamic pooling and a global model. The verified modality is text-to-text: potential future extensions to images or audio are not available capabilities. The repository creation date in August is not used as the release date.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | allenai/Bolmo-1B-Stage1 |
|---|---|
| Model type | Text and reasoning |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted with Transformers and xlstm |
| Release | 07 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text |
| Exit | Text |
| Context window | 4,096 internal positions; variable byte length |
| maximum output | max_new_tokens counts bytes; maximum not published |
| Reasoning | Base model; reasoning effort not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Transformers |
Research into tokenization, spelling, whitespace, unusual strings and dynamic compression. The profile identifies the exact checkpoint and training stage; the declared language is English. Bwen derives from Qwen3 8B Base and Blama from Llama 3 8B; Bolmo 1B and 7B start from Olmo. Do not confuse these base models with instruction-tuned assistants.
Stage 1 checkpoints keep the original global model frozen; they are not equivalent to their fully trained versions. The configuration declares internal model positions, not a guaranteed input-byte limit. max_new_tokens measures bytes, not the commercial tokens of other APIs. The official example uses trust_remote_code: review the code and pin a revision before running it. Models can produce harmful content, biases and incorrect claims without safety filtering.
03 / CAPABILITIES
What it can do
Orientation
Research into tokenization, spelling, whitespace, unusual strings and dynamic compression. The profile identifies the exact checkpoint and training stage; the declared language is English. Bwen derives from Qwen3 8B Base and Blama from Llama 3 8B; Bolmo 1B and 7B start from Olmo. Do not confuse these base models with instruction-tuned assistants.
Context
4,096 internal positions; variable byte length · max_new_tokens counts bytes; maximum not published
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
We have not verified a commercial endpoint or an official per-token rate. Numeric cost fields remain empty: downloadable weights do not mean free inference. Budget for memory, compute and maintenance. The documentation uses Python 3.11, Transformers 4.57.3 and xlstm 2.0.4. Bwen and Bolmo Stage 1 declare Apache 2.0; Blama retains Llama 3 Community License, with its own conditions. Also review Ai2’s responsible use guidelines.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Ai2 reports aggregate results and describes improvements in byte-level modeling, but we have not verified comparable numerical scores for this checkpoint. We do not transfer results between Bwen, Blama, Bolmo or between Stage 1 and full training. Our own evaluation is pending. For reproduction, retain revision, stage, data, compression, generated bytes, hardware and latency; normalize units before comparing with subword-token models.
i No verified numerical scores per checkpoint; Inferama evaluation pending.
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Bolmo 1B Stage 1
Stage 1: trained byte-level components; global model parameters unchanged.
ANALYSIS
Related analysis
07 / SOURCES