Характеристики

Лицензия: Своя лицензия
Базовая модель: OLMo-2-0425-1B-Instruct
Библиотека: transformers
Опубликована: 26 мар 2026

Ссылки

bolt-instruct-32b

Генерация текстаТекст → текст15 профессий

Англоязычная модель на базе OLMo для диалогов, выполнения инструкций, RAG, обработки документов и генерации кода с контекстом до 65 536 токенов.

66скачиваний за 30 дней0лайковвеса обновлены30 апр 2026

Бенчмарки

Как модель справилась с задачами пользователей ailist и что о ней публикуют открытые лидерборды

Публичные оценки

Из открытых лидербордов. Сравнивайте результаты одного теста при одинаковых условиях — отсутствие оценки не значит, что модель хуже.

В подключённых источниках результатов для этой модели пока нет.

Что умеет модель

Описание составлено по карточке модели и метаданным Hugging Face

Bolt Instruct 32B — англоязычная причинно-следственная языковая модель с настройкой на выполнение инструкций от AI Squared, основанная на объединении моделей Allen Institute for AI OLMo. Вариант 32B предназначен для диалогов, рассуждений и генерации высокого качества; карточка указывает OLMo-3.1-32B-Instruct среди базовых моделей и описывает обучение семейства методом SFT, а для моделей 7B и 32B — с применением QLoRA.

Модель подходит для чат-ассистентов и корпоративных сценариев: генерации структурированных ответов, извлечения сведений из документов, суммаризации, RAG и помощи с кодом. Она принимает текст и генерирует текст, поддерживает контекст до 65 536 токенов. В метаданных указана лицензия other; конкретные условия использования следует сверить с карточкой лицензии репозитория. Модель может галлюцинировать без опоры на источники, а качество зависит от размера варианта. Для задач с высокими последствиями требуется проверка человеком.

Сценарии применения

создание чат-ассистента для корпоративных пользователей
формирование JSON-ответов по заданной схеме
извлечение фактов из корпоративных документов
суммаризация длинных текстов
генерация ответов на основе найденных документов в RAG-системе
написание и преобразование кода

Ограничения

Модель может галлюцинировать без опоры на источники; качество зависит от размера варианта. Для применения в высокорисковых областях необходим контроль специалиста.

Для каких профессий подходит15

Разделы каталога профессий, в задачах которых используется эта модель

Все профессии

II.2Разработка и данныеСоздание и понимание программного кодаAI помогает писать, объяснять и связывать программный код с требованиями и существующей системой.20 задач21 443модели II.6Разработка и данныеДокументация и сопровождение разработкиAI помогает поддерживать понятную техническую документацию и ускорять передачу знаний.18 задач38 015моделей I.3Универсальные способностиИзвлечение и структурирование данныхAI помогает превращать разрозненные данные в структуру, пригодную для поиска, аналитики и автоматизации.18 задач25 791модель I.5Универсальные способностиСоздание и редактирование текстаAI помогает создавать и улучшать текстовые материалы, сохраняя смысл и требования автора.18 задач44 385моделей

Также подходит для

Карточка модели

README от автора на Hugging Face, как он его опубликовал

Открыть на Hugging Face

Bolt Instruct Models

Bolt Instruct is a family of instruction-tuned language models designed for high-quality generation, reasoning, and enterprise workflows.

These models are fine-tuned from Allen Institute for AI OLMo instruct models and optimized for:

General conversational AI
Structured and controllable generation
Retrieval-Augmented Generation (RAG)
Enterprise document understanding
Code generation and transformation

Model Overview

Bolt Instruct models provide strong instruction-following capabilities across diverse tasks with robust long-context support.

Key design goals:

Strong instruction adherence
High-quality structured outputs (JSON, extraction)
RAG-grounded responses
Long-context support (65k tokens for 7B and 32B)
Balanced chat, reasoning, and coding performance

Model Variants

Model	Base Model	Positioning
bolt-instruct-1b	allenai/OLMo-2-0425-1B-Instruct	Lightweight / low-latency
bolt-instruct-7b	allenai/OLMo-3-7B-Instruct	Balanced
bolt-instruct-32b	allenai/OLMo-3.1-32B-Instruct	Highest quality

Model Details

Type: Causal LM (instruction-tuned)
Max context: 65,536 tokens (7B and 32B), 4,096 tokens (1B)
Training context: 32k (7B), 16k (32B), 4k (1B)

Capabilities

Chat / multi-turn dialogue
Instruction following
Structured output (JSON)
Summarization & transformation
Extraction
RAG generation
Code generation

Training

Method: Supervised Fine-Tuning (SFT)
Dataset size: ~125k conversations
Eval set: ~10k examples
Data mix: public + synthetic + internal tasks

Training Approach

1B → full fine-tune
7B / 32B → QLoRA (4-bit)

Hardware

1× A100 80GB GPU

Intended Use

Chat assistants
Enterprise copilots
RAG pipelines
Document processing
Structured extraction
Code assistance

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "aisquared/bolt-instruct-7b"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

Evaluation

To evaluate these models, we ran a subset of tasks using the Eleuther AI Language Model Evaluation Harness. Below are the metrics for each model.

Language Model Evaluation Harness

Evaluation results for aisquared/bolt-instruct-1b:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
arc_challenge	1	none	0	acc	↑	0.3490	±	0.0139
		none	0	acc_norm	↑	0.3823	±	0.0142
arc_easy	1	none	0	acc	↑	0.6098	±	0.0100
		none	0	acc_norm	↑	0.5560	±	0.0102
bbh	3	get-answer		exact_match	↑	0.3081	±	0.0052
- bbh_cot_fewshot_boolean_expressions	4	get-answer	3	exact_match	↑	0.5840	±	0.0312
- bbh_cot_fewshot_causal_judgement	4	get-answer	3	exact_match	↑	0.5508	±	0.0365
- bbh_cot_fewshot_date_understanding	4	get-answer	3	exact_match	↑	0.2600	±	0.0278
- bbh_cot_fewshot_disambiguation_qa	4	get-answer	3	exact_match	↑	0.3640	±	0.0305
- bbh_cot_fewshot_dyck_languages	4	get-answer	3	exact_match	↑	0.0040	±	0.0040
- bbh_cot_fewshot_formal_fallacies	4	get-answer	3	exact_match	↑	0.5040	±	0.0317
- bbh_cot_fewshot_geometric_shapes	4	get-answer	3	exact_match	↑	0.0920	±	0.0183
- bbh_cot_fewshot_hyperbaton	4	get-answer	3	exact_match	↑	0.5240	±	0.0316
- bbh_cot_fewshot_logical_deduction_five_objects	4	get-answer	3	exact_match	↑	0.1720	±	0.0239
- bbh_cot_fewshot_logical_deduction_seven_objects	4	get-answer	3	exact_match	↑	0.1080	±	0.0197
- bbh_cot_fewshot_logical_deduction_three_objects	4	get-answer	3	exact_match	↑	0.3520	±	0.0303
- bbh_cot_fewshot_movie_recommendation	4	get-answer	3	exact_match	↑	0.5040	±	0.0317
- bbh_cot_fewshot_multistep_arithmetic_two	4	get-answer	3	exact_match	↑	0.0600	±	0.0151
- bbh_cot_fewshot_navigate	4	get-answer	3	exact_match	↑	0.5560	±	0.0315
- bbh_cot_fewshot_object_counting	4	get-answer	3	exact_match	↑	0.4360	±	0.0314
- bbh_cot_fewshot_penguins_in_a_table	4	get-answer	3	exact_match	↑	0.2123	±	0.0340
- bbh_cot_fewshot_reasoning_about_colored_objects	4	get-answer	3	exact_match	↑	0.2440	±	0.0272
- bbh_cot_fewshot_ruin_names	4	get-answer	3	exact_match	↑	0.2440	±	0.0272
- bbh_cot_fewshot_salient_translation_error_detection	4	get-answer	3	exact_match	↑	0.1920	±	0.0250
- bbh_cot_fewshot_snarks	4	get-answer	3	exact_match	↑	0.3989	±	0.0368
- bbh_cot_fewshot_sports_understanding	4	get-answer	3	exact_match	↑	0.6560	±	0.0301
- bbh_cot_fewshot_temporal_sequences	4	get-answer	3	exact_match	↑	0.2760	±	0.0283
- bbh_cot_fewshot_tracking_shuffled_objects_five_objects	4	get-answer	3	exact_match	↑	0.1920	±	0.0250
- bbh_cot_fewshot_tracking_shuffled_objects_seven_objects	4	get-answer	3	exact_match	↑	0.0360	±	0.0118
- bbh_cot_fewshot_tracking_shuffled_objects_three_objects	4	get-answer	3	exact_match	↑	0.2840	±	0.0286
- bbh_cot_fewshot_web_of_lies	4	get-answer	3	exact_match	↑	0.5240	±	0.0316
- bbh_cot_fewshot_word_sorting	4	get-answer	3	exact_match	↑	0.0360	±	0.0118
gsm8k	3	flexible-extract	5	exact_match	↑	0.5072	±	0.0138
		strict-match	5	exact_match	↑	0.4943	±	0.0138
hellaswag	1	none	0	acc	↑	0.4729	±	0.0050
		none	0	acc_norm	↑	0.6181	±	0.0048
mmlu_pro	2	custom-extract		exact_match	↑	0.1435	±	0.0032
- biology	3	custom-extract	5	exact_match	↑	0.2050	±	0.0151
- business	3	custom-extract	5	exact_match	↑	0.1369	±	0.0122
- chemistry	3	custom-extract	5	exact_match	↑	0.0848	±	0.0083
- computer_science	3	custom-extract	5	exact_match	↑	0.1415	±	0.0172
- economics	3	custom-extract	5	exact_match	↑	0.1943	±	0.0136
- engineering	3	custom-extract	5	exact_match	↑	0.0929	±	0.0093
- health	3	custom-extract	5	exact_match	↑	0.1528	±	0.0126
- history	3	custom-extract	5	exact_match	↑	0.1549	±	0.0186
- law	3	custom-extract	5	exact_match	↑	0.1081	±	0.0094
- math	3	custom-extract	5	exact_match	↑	0.1414	±	0.0095
- other	3	custom-extract	5	exact_match	↑	0.1916	±	0.0130
- philosophy	3	custom-extract	5	exact_match	↑	0.1383	±	0.0155
- physics	3	custom-extract	5	exact_match	↑	0.1186	±	0.0090
- psychology	3	custom-extract	5	exact_match	↑	0.2130	±	0.0145
truthfulqa_mc2	3	none	0	acc	↑	0.4734	±	0.0153
winogrande	1	none	0	acc	↑	0.6156	±	0.0137

Evaluation results for aisquared/bolt-instruct-7b:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
arc_challenge	1	none	0	acc	↑	0.4778	±	0.0146
		none	0	acc_norm	↑	0.4957	±	0.0146
arc_easy	1	none	0	acc	↑	0.7534	±	0.0088
		none	0	acc_norm	↑	0.7311	±	0.0091
bbh	3	get-answer		exact_match	↑	0.3038	±	0.0047
- bbh_cot_fewshot_boolean_expressions	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_causal_judgement	4	get-answer	3	exact_match	↑	0.5668	±	0.0363
- bbh_cot_fewshot_date_understanding	4	get-answer	3	exact_match	↑	0.4480	±	0.0315
- bbh_cot_fewshot_disambiguation_qa	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_dyck_languages	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_formal_fallacies	4	get-answer	3	exact_match	↑	0.2240	±	0.0264
- bbh_cot_fewshot_geometric_shapes	4	get-answer	3	exact_match	↑	0.2960	±	0.0289
- bbh_cot_fewshot_hyperbaton	4	get-answer	3	exact_match	↑	0.5200	±	0.0317
- bbh_cot_fewshot_logical_deduction_five_objects	4	get-answer	3	exact_match	↑	0.0200	±	0.0089
- bbh_cot_fewshot_logical_deduction_seven_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_logical_deduction_three_objects	4	get-answer	3	exact_match	↑	0.6720	±	0.0298
- bbh_cot_fewshot_movie_recommendation	4	get-answer	3	exact_match	↑	0.1200	±	0.0206
- bbh_cot_fewshot_multistep_arithmetic_two	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_navigate	4	get-answer	3	exact_match	↑	0.5560	±	0.0315
- bbh_cot_fewshot_object_counting	4	get-answer	3	exact_match	↑	0.1520	±	0.0228
- bbh_cot_fewshot_penguins_in_a_table	4	get-answer	3	exact_match	↑	0.4110	±	0.0409
- bbh_cot_fewshot_reasoning_about_colored_objects	4	get-answer	3	exact_match	↑	0.1880	±	0.0248
- bbh_cot_fewshot_ruin_names	4	get-answer	3	exact_match	↑	0.4800	±	0.0317
- bbh_cot_fewshot_salient_translation_error_detection	4	get-answer	3	exact_match	↑	0.4760	±	0.0316
- bbh_cot_fewshot_snarks	4	get-answer	3	exact_match	↑	0.2921	±	0.0342
- bbh_cot_fewshot_sports_understanding	4	get-answer	3	exact_match	↑	0.6760	±	0.0297
- bbh_cot_fewshot_temporal_sequences	4	get-answer	3	exact_match	↑	0.5880	±	0.0312
- bbh_cot_fewshot_tracking_shuffled_objects_five_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_tracking_shuffled_objects_seven_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_tracking_shuffled_objects_three_objects	4	get-answer	3	exact_match	↑	0.8280	±	0.0239
- bbh_cot_fewshot_web_of_lies	4	get-answer	3	exact_match	↑	0.6560	±	0.0301
- bbh_cot_fewshot_word_sorting	4	get-answer	3	exact_match	↑	0.1400	±	0.0220
gsm8k	3	flexible-extract	5	exact_match	↑	0.7998	±	0.0110
		strict-match	5	exact_match	↑	0.7392	±	0.0121
hellaswag	1	none	0	acc	↑	0.4882	±	0.0050
		none	0	acc_norm	↑	0.6165	±	0.0049
mmlu_pro	2	custom-extract		exact_match	↑	0.4978	±	0.0044
- biology	3	custom-extract	5	exact_match	↑	0.6848	±	0.0174
- business	3	custom-extract	5	exact_match	↑	0.5729	±	0.0176
- chemistry	3	custom-extract	5	exact_match	↑	0.5380	±	0.0148
- computer_science	3	custom-extract	5	exact_match	↑	0.5878	±	0.0243
- economics	3	custom-extract	5	exact_match	↑	0.5592	±	0.0171
- engineering	3	custom-extract	5	exact_match	↑	0.2405	±	0.0137
- health	3	custom-extract	5	exact_match	↑	0.4670	±	0.0175
- history	3	custom-extract	5	exact_match	↑	0.3727	±	0.0248
- law	3	custom-extract	5	exact_match	↑	0.2525	±	0.0131
- math	3	custom-extract	5	exact_match	↑	0.7158	±	0.0123
- other	3	custom-extract	5	exact_match	↑	0.4351	±	0.0163
- philosophy	3	custom-extract	5	exact_match	↑	0.4128	±	0.0221
- physics	3	custom-extract	5	exact_match	↑	0.5142	±	0.0139
- psychology	3	custom-extract	5	exact_match	↑	0.5602	±	0.0176
truthfulqa_mc2	3	none	0	acc	↑	0.5666	±	0.0162
winogrande	1	none	0	acc	↑	0.6385	±	0.0135

Evaluation results for aisquared/bolt-instruct-32b:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
arc_challenge	1	none	0	acc	↑	0.5776	±	0.0144
		none	0	acc_norm	↑	0.6007	±	0.0143
arc_easy	1	none	0	acc	↑	0.8333	±	0.0076
		none	0	acc_norm	↑	0.8228	±	0.0078
bbh	3	get-answer		exact_match	↑	0.3087	±	0.0048
- bbh_cot_fewshot_boolean_expressions	4	get-answer	3	exact_match	↑	0.5760	±	0.0313
- bbh_cot_fewshot_causal_judgement	4	get-answer	3	exact_match	↑	0.5882	±	0.0361
- bbh_cot_fewshot_date_understanding	4	get-answer	3	exact_match	↑	0.6640	±	0.0299
- bbh_cot_fewshot_disambiguation_qa	4	get-answer	3	exact_match	↑	0.1920	±	0.0250
- bbh_cot_fewshot_dyck_languages	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_formal_fallacies	4	get-answer	3	exact_match	↑	0.0480	±	0.0135
- bbh_cot_fewshot_geometric_shapes	4	get-answer	3	exact_match	↑	0.2760	±	0.0283
- bbh_cot_fewshot_hyperbaton	4	get-answer	3	exact_match	↑	0.3200	±	0.0296
- bbh_cot_fewshot_logical_deduction_five_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_logical_deduction_seven_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_logical_deduction_three_objects	4	get-answer	3	exact_match	↑	0.5400	±	0.0316
- bbh_cot_fewshot_movie_recommendation	4	get-answer	3	exact_match	↑	0.6000	±	0.0310
- bbh_cot_fewshot_multistep_arithmetic_two	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_navigate	4	get-answer	3	exact_match	↑	0.0160	±	0.0080
- bbh_cot_fewshot_object_counting	4	get-answer	3	exact_match	↑	0.5120	±	0.0317
- bbh_cot_fewshot_penguins_in_a_table	4	get-answer	3	exact_match	↑	0.2945	±	0.0379
- bbh_cot_fewshot_reasoning_about_colored_objects	4	get-answer	3	exact_match	↑	0.2280	±	0.0266
- bbh_cot_fewshot_ruin_names	4	get-answer	3	exact_match	↑	0.5120	±	0.0317
- bbh_cot_fewshot_salient_translation_error_detection	4	get-answer	3	exact_match	↑	0.5440	±	0.0316
- bbh_cot_fewshot_snarks	4	get-answer	3	exact_match	↑	0.7079	±	0.0342
- bbh_cot_fewshot_sports_understanding	4	get-answer	3	exact_match	↑	0.4880	±	0.0317
- bbh_cot_fewshot_temporal_sequences	4	get-answer	3	exact_match	↑	0.3120	±	0.0294
- bbh_cot_fewshot_tracking_shuffled_objects_five_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_tracking_shuffled_objects_seven_objects	4	get-answer	3	exact_match	↑	0.0000	±	0.0000
- bbh_cot_fewshot_tracking_shuffled_objects_three_objects	4	get-answer	3	exact_match	↑	0.6280	±	0.0306
- bbh_cot_fewshot_web_of_lies	4	get-answer	3	exact_match	↑	0.4400	±	0.0315
- bbh_cot_fewshot_word_sorting	4	get-answer	3	exact_match	↑	0.0280	±	0.0105
gsm8k	3	flexible-extract	5	exact_match	↑	0.8795	±	0.0090
		strict-match	5	exact_match	↑	0.7801	±	0.0114
hellaswag	1	none	0	acc	↑	0.5407	±	0.0050
		none	0	acc_norm	↑	0.6763	±	0.0047
mmlu_pro	2	custom-extract		exact_match	↑	0.6340	±	0.0042
- biology	3	custom-extract	5	exact_match	↑	0.8117	±	0.0146
- business	3	custom-extract	5	exact_match	↑	0.6907	±	0.0165
- chemistry	3	custom-extract	5	exact_match	↑	0.6431	±	0.0142
- computer_science	3	custom-extract	5	exact_match	↑	0.6951	±	0.0228
- economics	3	custom-extract	5	exact_match	↑	0.7405	±	0.0151
- engineering	3	custom-extract	5	exact_match	↑	0.3447	±	0.0153
- health	3	custom-extract	5	exact_match	↑	0.6540	±	0.0166
- history	3	custom-extract	5	exact_match	↑	0.5512	±	0.0255
- law	3	custom-extract	5	exact_match	↑	0.3860	±	0.0147
- math	3	custom-extract	5	exact_match	↑	0.7979	±	0.0109
- other	3	custom-extract	5	exact_match	↑	0.6028	±	0.0161
- philosophy	3	custom-extract	5	exact_match	↑	0.5912	±	0.0220
- physics	3	custom-extract	5	exact_match	↑	0.6551	±	0.0132
- psychology	3	custom-extract	5	exact_match	↑	0.7243	±	0.0158
truthfulqa_mc2	3	none	0	acc	↑	0.6906	±	0.0153
winogrande	1	none	0	acc	↑	0.6630	±	0.0133

Limitations

May hallucinate without grounding
Performance varies by model size
Not suitable for high-risk domains without oversight

License

Bolt Instruct is released under the AI Squared Community License.

Как запустить

На своей GPU в ailist. Цена — по текущим предложениям провайдеров

GPU в ailist

vLLM

Нельзя запустить

Доступ к модели ограничен автором; у платформы нет доступа к весам.

Ревизия весов: d595b79