BOLMO: Byte Level Transformer (that you can use today)
* no tokenizer, completely gone, it directly interprets bytes
* performance on par with OLMo 3
* everything is open, as usual for @ai2.bsky.social
allenai.org/blog/bolmo
* no tokenizer, completely gone, it directly interprets bytes
* performance on par with OLMo 3
* everything is open, as usual for @ai2.bsky.social
allenai.org/blog/bolmo
Introducing Bolmo: Byteifying the next generation of language models | Ai2
Bolmo is new a byte-level family built by adapting Olmo 3 into a fast, flexible byte-based model with a short extra training run.
allenai.org
December 16, 2025 at 3:27 PM
BOLMO: Byte Level Transformer (that you can use today)
* no tokenizer, completely gone, it directly interprets bytes
* performance on par with OLMo 3
* everything is open, as usual for @ai2.bsky.social
allenai.org/blog/bolmo
* no tokenizer, completely gone, it directly interprets bytes
* performance on par with OLMo 3
* everything is open, as usual for @ai2.bsky.social
allenai.org/blog/bolmo
ok cool EXCEPT for the model names
wdym you have olmo, molmo, and fucking bolmo
wdym you have olmo, molmo, and fucking bolmo
haven't heard much about ai2 before they released olmo 3, but wow their stuff is both really interesting and really cool (its all open, not just the weights but the training steps and datasets too!)
December 16, 2025 at 12:53 AM
ok cool EXCEPT for the model names
wdym you have olmo, molmo, and fucking bolmo
wdym you have olmo, molmo, and fucking bolmo
uh oh: "By operating directly over raw UTF-8 bytes..."
December 15, 2025 at 5:22 PM
uh oh: "By operating directly over raw UTF-8 bytes..."
We're excited about what's next, including scaling byteifying to larger models.
📝 Blog: buff.ly/WjU6bXV
⬇️ Download Bolmo 7B: buff.ly/pkCdFWl | 1B: buff.ly/QRBm6hC
📄 Report: buff.ly/08ouIoy
📝 Blog: buff.ly/WjU6bXV
⬇️ Download Bolmo 7B: buff.ly/pkCdFWl | 1B: buff.ly/QRBm6hC
📄 Report: buff.ly/08ouIoy
Introducing Bolmo: Byteifying the next generation of language models | Ai2
Bolmo is new a byte-level family built by adapting Olmo 3 into a fast, flexible byte-based model with a short extra training run.
allenai.org
December 15, 2025 at 5:19 PM
We're excited about what's next, including scaling byteifying to larger models.
📝 Blog: buff.ly/WjU6bXV
⬇️ Download Bolmo 7B: buff.ly/pkCdFWl | 1B: buff.ly/QRBm6hC
📄 Report: buff.ly/08ouIoy
📝 Blog: buff.ly/WjU6bXV
⬇️ Download Bolmo 7B: buff.ly/pkCdFWl | 1B: buff.ly/QRBm6hC
📄 Report: buff.ly/08ouIoy
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3—and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵
December 15, 2025 at 5:19 PM
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3—and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵
Fun with tokenization - not sure why this happens exactly
Claude has no token for Bolmo (because Bolmo is new) and so can't directly encode that name, but it can *decode* it somehow - it might get split into multiple tokens in some sentences, but not others?
also extremely ironic in this context
Claude has no token for Bolmo (because Bolmo is new) and so can't directly encode that name, but it can *decode* it somehow - it might get split into multiple tokens in some sentences, but not others?
also extremely ironic in this context
January 29, 2026 at 5:31 PM
Fun with tokenization - not sure why this happens exactly
Claude has no token for Bolmo (because Bolmo is new) and so can't directly encode that name, but it can *decode* it somehow - it might get split into multiple tokens in some sentences, but not others?
also extremely ironic in this context
Claude has no token for Bolmo (because Bolmo is new) and so can't directly encode that name, but it can *decode* it somehow - it might get split into multiple tokens in some sentences, but not others?
also extremely ironic in this context
Bolmo takes an existing Olmo 3 7B checkpoint and retrofits it into a fast, flexible byte-level architecture. It skips hand-engineered vocabularies and operates directly on UTF-8 bytes, handling spelling, edge cases, & multilingual scripts naturally.
December 15, 2025 at 5:19 PM
Bolmo takes an existing Olmo 3 7B checkpoint and retrofits it into a fast, flexible byte-level architecture. It skips hand-engineered vocabularies and operates directly on UTF-8 bytes, handling spelling, edge cases, & multilingual scripts naturally.
The Allen Institute of AI launches Bolmo 7B and Bolmo 1B, claiming they are "the first fully open byte-level language models", built on its Olmo 3 models (Emilia David/VentureBeat)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
December 16, 2025 at 1:05 AM
The Allen Institute of AI launches Bolmo 7B and Bolmo 1B, claiming they are "the first fully open byte-level language models", built on its Olmo 3 models (Emilia David/VentureBeat)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
We keep Olmo 3's original backbone & capabilities, adding a lightweight byte stack so Bolmo can reason over bytes without discarding prior work. The result: a byte-level model with Olmo 3's strengths + finer-grained text understanding. 🚀
December 15, 2025 at 5:19 PM
We keep Olmo 3's original backbone & capabilities, adding a lightweight byte stack so Bolmo can reason over bytes without discarding prior work. The result: a byte-level model with Olmo 3's strengths + finer-grained text understanding. 🚀
On our eval suite & character-focused benchmarks like CUTE & EXECUTE, Bolmo matches/surpasses subword models while excelling at character-level reasoning. Once you byteify a base model, you can import capabilities from post-trained checkpoints via weight arithmetic.
December 15, 2025 at 5:19 PM
On our eval suite & character-focused benchmarks like CUTE & EXECUTE, Bolmo matches/surpasses subword models while excelling at character-level reasoning. Once you byteify a base model, you can import capabilities from post-trained checkpoints via weight arithmetic.
ask them to draw an illust of their family
i wanna see what bolmo looks like
i wanna see what bolmo looks like
December 16, 2025 at 1:07 AM
ask them to draw an illust of their family
i wanna see what bolmo looks like
i wanna see what bolmo looks like
the key: bootstrap from OLMo3
December 16, 2025 at 3:27 PM
the key: bootstrap from OLMo3
winbuzzer.com/2025/12/16/a...
Byteification: AI2's New Bolmo AI Model Cuts AI Training Costs by 99%
#AI #AI2 #LLMs #OpenSourceAI #AIResearch #MachineLearning #Bolmo #ByteLevelAI #Tokenization #ModelEfficiency #DeepLearning
Byteification: AI2's New Bolmo AI Model Cuts AI Training Costs by 99%
#AI #AI2 #LLMs #OpenSourceAI #AIResearch #MachineLearning #Bolmo #ByteLevelAI #Tokenization #ModelEfficiency #DeepLearning
Byteification: AI2's New Bolmo AI Model Cuts AI Training Costs by 99% - WinBuzzer
AI2 has unveiled Bolmo, a byte-level model created by retrofitting its OLMo 3 model with <1% of the compute budget.
winbuzzer.com
December 16, 2025 at 10:16 AM
winbuzzer.com/2025/12/16/a...
Byteification: AI2's New Bolmo AI Model Cuts AI Training Costs by 99%
#AI #AI2 #LLMs #OpenSourceAI #AIResearch #MachineLearning #Bolmo #ByteLevelAI #Tokenization #ModelEfficiency #DeepLearning
Byteification: AI2's New Bolmo AI Model Cuts AI Training Costs by 99%
#AI #AI2 #LLMs #OpenSourceAI #AIResearch #MachineLearning #Bolmo #ByteLevelAI #Tokenization #ModelEfficiency #DeepLearning
Архитектура Bolmo позволяет эффективно обучать языковые модели на уровне байтов без ущерба для качества.
Институт искусственного интеллекта Аллена (AI2) разработал Bolmo, новое семейство языковых моделей байтового уровня. Bolmo использует существующие модели Olmo 3 …
Telegram ИИ Дайджест
#ai #news
Институт искусственного интеллекта Аллена (AI2) разработал Bolmo, новое семейство языковых моделей байтового уровня. Bolmo использует существующие модели Olmo 3 …
Telegram ИИ Дайджест
#ai #news
Bolmo’s architecture unlocks efficient byte‑level LM training without sacrificing quality
venturebeat.com
December 16, 2025 at 12:16 AM
Архитектура Bolmo позволяет эффективно обучать языковые модели на уровне байтов без ущерба для качества.
Институт искусственного интеллекта Аллена (AI2) разработал Bolmo, новое семейство языковых моделей байтового уровня. Bolmo использует существующие модели Olmo 3 …
Telegram ИИ Дайджест
#ai #news
Институт искусственного интеллекта Аллена (AI2) разработал Bolmo, новое семейство языковых моделей байтового уровня. Bolmo использует существующие модели Olmo 3 …
Telegram ИИ Дайджест
#ai #news
[2025-12-18] 📚 Updates in #MoE
(1) Improving Recursive Transformers with Mixture of LoRAs
(2) <a href="https://researchtrend.ai/papers/2512.15586" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">Bolmo: Byteifying the Next Generation of Language Models
(3) Bolmo: Byteifying the Next Generation of Language Models
🔍 More at researchtrend.ai/communities/MoE
(1) Improving Recursive Transformers with Mixture of LoRAs
(2) <a href="https://researchtrend.ai/papers/2512.15586" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">Bolmo: Byteifying the Next Generation of Language Models
(3) Bolmo: Byteifying the Next Generation of Language Models
🔍 More at researchtrend.ai/communities/MoE
December 18, 2025 at 3:07 AM
[2025-12-18] 📚 Updates in #MoE
(1) Improving Recursive Transformers with Mixture of LoRAs
(2) <a href="https://researchtrend.ai/papers/2512.15586" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">Bolmo: Byteifying the Next Generation of Language Models
(3) Bolmo: Byteifying the Next Generation of Language Models
🔍 More at researchtrend.ai/communities/MoE
(1) Improving Recursive Transformers with Mixture of LoRAs
(2) <a href="https://researchtrend.ai/papers/2512.15586" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link" target="_blank" rel="noopener" data-link="bsky">Bolmo: Byteifying the Next Generation of Language Models
(3) Bolmo: Byteifying the Next Generation of Language Models
🔍 More at researchtrend.ai/communities/MoE
Ah, yes, as I said...
We'll wait.
You can have some reading on a strategy that uses utf8 instead of "words". Plenty to look up from here including some that can indeed make images.
allenai.org/blog/bolmo
We'll wait.
You can have some reading on a strategy that uses utf8 instead of "words". Plenty to look up from here including some that can indeed make images.
allenai.org/blog/bolmo
Introducing Bolmo: Byteifying the next generation of language models | Ai2
Bolmo is new a byte-level family built by adapting Olmo 3 into a fast, flexible byte-based model with a short extra training run.
allenai.org
December 22, 2025 at 3:48 AM
Ah, yes, as I said...
We'll wait.
You can have some reading on a strategy that uses utf8 instead of "words". Plenty to look up from here including some that can indeed make images.
allenai.org/blog/bolmo
We'll wait.
You can have some reading on a strategy that uses utf8 instead of "words". Plenty to look up from here including some that can indeed make images.
allenai.org/blog/bolmo
Feed: "Techmeme"
Published on Tuesday, December 16, 2025
Published on Tuesday, December 16, 2025
Allen Institute for AI launches Bolmo 7B and Bolmo 1B, claiming they are "the first fully open byte-level language models", built on its Olmo 3 models (Emilia David/VentureBeat)
By Emilia David / VentureBeat. View the full context on Techmeme.
www.techmeme.com
December 16, 2025 at 8:50 AM
Feed: "Techmeme"
Published on Tuesday, December 16, 2025
Published on Tuesday, December 16, 2025
Bolmo’s new architecture lets you train LMs at the byte level—no more tokenizer headaches, faster runs, and true multilingual support. Curious how it stacks up against Dolma 3 or Olmo? Dive in! #Bolmo #ByteLevel #Multilingual
🔗 aidailypost.com/news/bolmo-a...
🔗 aidailypost.com/news/bolmo-a...
December 15, 2025 at 10:42 PM
Bolmo’s new architecture lets you train LMs at the byte level—no more tokenizer headaches, faster runs, and true multilingual support. Curious how it stacks up against Dolma 3 or Olmo? Dive in! #Bolmo #ByteLevel #Multilingual
🔗 aidailypost.com/news/bolmo-a...
🔗 aidailypost.com/news/bolmo-a...
Introducing Bolmo: Byteifying the next generation of language models https://lobste.rs/s/wuahgz ##ai
Introducing Bolmo: Byteifying the next generation of language models | Ai2
Bolmo is new a byte-level family built by adapting Olmo 3 into a fast, flexible byte-based model with a short extra training run.
allenai.org
December 15, 2025 at 5:30 PM
Introducing Bolmo: Byteifying the next generation of language models https://lobste.rs/s/wuahgz ##ai
This tweet appeared under this Techmeme headline:
@allen_ai:
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3--and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵 [image]
@allen_ai:
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3--and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵 [image]
December 16, 2025 at 1:08 AM
This tweet appeared under this Techmeme headline:
@allen_ai:
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3--and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵 [image]
@allen_ai:
Introducing Bolmo, a new family of byte-level language models built by "byteifying" our open Olmo 3--and to our knowledge, the first fully open byte-level LM to match or surpass SOTA subword models across a wide range of tasks. 🧵 [image]
Very interesting. Dropping tokenization inspires more multimodal capabilities for future developments.
https://allenai.org/blog/bolmo
https://allenai.org/blog/bolmo
December 16, 2025 at 4:19 PM
Very interesting. Dropping tokenization inspires more multimodal capabilities for future developments.
https://allenai.org/blog/bolmo
https://allenai.org/blog/bolmo
Benjamin Minixhofer, Tyler Murray, Tomasz Limisiewicz, Anna Korhonen, Luke Zettlemoyer, Noah A. Smith, Edoardo M. Ponti, Luca Soldaini, Valentin Hofmann
Bolmo: Byteifying the Next Generation of Language Models
https://arxiv.org/abs/2512.15586
Bolmo: Byteifying the Next Generation of Language Models
https://arxiv.org/abs/2512.15586
December 18, 2025 at 6:22 AM
Benjamin Minixhofer, Tyler Murray, Tomasz Limisiewicz, Anna Korhonen, Luke Zettlemoyer, Noah A. Smith, Edoardo M. Ponti, Luca Soldaini, Valentin Hofmann
Bolmo: Byteifying the Next Generation of Language Models
https://arxiv.org/abs/2512.15586
Bolmo: Byteifying the Next Generation of Language Models
https://arxiv.org/abs/2512.15586
🌍 Developers & researchers, say goodbye to tokenization woes! 🚀 Ai2's Bolmo models revolutionize multilingual processing with unmatched performance and no vocab limits. 💪✨ Discover how they can transform your projects! #AI #MultilingualModels #Bolmo LINK
December 16, 2025 at 8:18 AM
🌍 Developers & researchers, say goodbye to tokenization woes! 🚀 Ai2's Bolmo models revolutionize multilingual processing with unmatched performance and no vocab limits. 💪✨ Discover how they can transform your projects! #AI #MultilingualModels #Bolmo LINK