GammaAGI/ai21-jamba-3b-merge-v1

GAMMAai Jamba 3B Merge v1

This repository contains a complete Jamba-family research checkpoint from the first GAMMAai fusion batch. It combines the compact instruction profile of ai21labs/AI21-Jamba2-3B with the reasoning-oriented Jamba 3B branch while retaining a single deployable JambaForCausalLM checkpoint.

Checkpoint

• Architecture: hybrid Mamba/attention JambaForCausalLM
• Hidden size: 2,560
• Layers: 28
• Attention cadence: one attention layer every 14 layers
• Precision: FP16
• Declared context window: 262,144 tokens
• Weight package: one complete model.safetensors file
• Tokenizer, chat template and configuration: included

The two parent checkpoints share the same Jamba 3B topology and Apache-2.0 license family. The public artifact is structurally complete. The exact coefficient history from the early fusion run was not preserved in this repository, so this release is presented as a research checkpoint rather than as a benchmark-certified replacement for either parent.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "GammaAGI/ai21-jamba-3b-merge-v1"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain why staged verification reduces model-training risk."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)
output = model.generate(inputs, max_new_tokens=160, do_sample=False)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

Intended Use

The checkpoint is intended for research, comparative inference, chromosome extraction and controlled downstream adaptation. Validate task quality, long-context behavior and safety on your own workload before production use.

Limitations

• No public benchmark superiority claim is made.
• The 262K configuration value is not a substitute for measured long-context quality at every length.
• Fusion can inherit limitations and biases from both parent checkpoints.
• Production deployment requires an independent generation and safety evaluation.

Project

GAMMAai develops verifier-first, staged model evolution through EvoStream, SafeTrain and measurable no-regress gates. This checkpoint is retained as a complete historical model product; temporary adapters, failed experiments and partial weight artifacts are maintained outside the public model catalog.