Tasks and inputs
Llama-3-Blama-8B is listed for Fill-mask, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Llama-3-Blama-8B is Ai2's byte-level retrofit of Llama 3 8B, released to show the Bolmo byteification method works beyond Ai2's own Olmo models.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Llama-3-Blama-8B exists to prove a point. Ai2's Bolmo method converts a conventional subword language model into one that reads and writes raw bytes, and until recently it had only been demonstrated on Ai2's own Olmo models. Retrofitting Meta's Llama 3 8B, a model trained by somebody else with a different recipe, shows the technique transfers. Ai2 released it on 7 October 2026 alongside a Qwen-based twin, when the Bolmo research was published in Nature. Ai2 also calls it simply "Blama 8B".
Because the tokeniser is a fixed vocabulary, every subword model inherits its blind spots: awkward splits of rare words, inconsistent whitespace, weak handling of unfamiliar scripts and exact strings. A byte-level model sees every character. Training one from scratch costs a fortune. Byteification gets there with a comparatively short additional run, keeping most of what the original network learned.
The catch is quality. Byteified checkpoints come close to their source models but do not fully match them, and the research community is still working out the trade-off between byte-level efficiency and character-level understanding.
| Llama-3-Blama-8B | Bwen 8B | |
|---|---|---|
| Source model | Llama 3 8B | Qwen3-8B |
| Licence | Llama 3 licence | Apache 2.0 |
| Released | 7 October 2026 | 7 October 2026 |
| Ai2's verdict | Close to its source | Strongest byteified model so far |
The licence is the practical difference. Blama inherits Meta's Llama 3 terms (tagged llama3), not Apache 2.0, so the usual Llama 3 conditions on use and attribution apply. If you want permissive terms and the stronger scores, the Qwen-based Bwen 8B is the better default. Blama is more interesting if your existing stack, prompts or fine-tuning data are built around Llama 3.
As with the other retrofits, Ai2 published Stage 1 versions in which the original Llama network stays frozen while the new byte components train. They are a cheaper starting point if you want to run your own byteification experiments rather than use the finished model.
allenai/Llama-3-Blama-8B, part of Ai2's Bolmo collection.trust_remote_code) and counts generation length in bytes, and you should assume nothing about this one until you have checked.The cleanest experiment is a pair comparison: the original Llama 3 8B against the byte version, on the same prompts. Use one set that stresses characters (misspellings, base64 strings, transliteration) and one ordinary set (reading comprehension, short QA). The first shows what you gain, the second what you lose. For context on the wider Llama lineage see Llama 3.1 405B and the Meta Llama provider page; Ai2's other work is on the Ai2 provider page.
Compare the model on the work you need to complete.
Llama-3-Blama-8B is listed for Fill-mask, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under llama3. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Microsoft
Aion 1.0 Instruct 2.3B is Microsoft's MIT-licensed on-device small language model and Phi Silica successor for Windows, with a 32K context.
Gemini 4 Neon is Google's mid-tier Gemini 4 model for complex coding and analysis, sitting between Hydrogen and Carbon at $3/$15 per 1M tokens.
OpenAI
GPT-6.2 Astra is OpenAI's November 2026 flagship for hard professional, scientific and agentic work, at $8/$40 per million tokens with a 1.05M context.
OpenAI
GPT-6.2 Sol is OpenAI's everyday model for coding, agents and writing: $2/$10 per million tokens, 1.05M context, released 17 November 2026.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.