---
language:
  - en
license: apache-2.0
tags:
  - knot
  - mamba2
  - affectively
  - edgework
  - aether
  - distributed-inference
  - edge-deployment
base_model: state-spaces/mamba-2.8b-hf
base_model_relation: quantized
pipeline_tag: text-generation
---

# Mamba 2.8b (KNOT)
> **Production-ready** KNOT (sovereign-format) mirror of [state-spaces/mamba-2.8b-hf](https://huggingface.co/state-spaces/mamba-2.8b-hf) for distributed text generation and conversation — powered by the [Aether](https://github.com/forkjoin-ai/aether) edge inference runtime on [Edgework.ai](https://edgework.ai).

## Model Details

| Property | Value |
|----------|-------|
| Base model | [state-spaces/mamba-2.8b-hf](https://huggingface.co/state-spaces/mamba-2.8b-hf) |
| Parameters | 2.8B |
| Architecture | Mamba |
| Quantization | — (lossless container) |
| Format | KNOT |
| Size | ~11.2 GB |
| License | apache-2.0 |

## Also available: `.knot` (sovereign format)

This repo ships [`mamba-2.8b.knot`](./mamba-2.8b.knot) — the model weights in the **KNOT** container that the [Aether](https://github.com/forkjoin-ai/aether) distributed-inference runtime loads natively (the GGUF, when present, sits right beside it). A KNOT is a single self-describing file with a JSON table-of-contents, so **any single tensor is one HTTP `Range` request** — ideal for streaming weights to edge nodes.

| | GGUF | KNOT |
|---|---|---|
| Container | format-specific header | single file, JSON table-of-contents |
| Per-tensor fetch | whole-file oriented | one tensor = one Range request |
| Ecosystem | broad (llama.cpp, …) | Aether / Gnosis runtime |

```bash
huggingface-cli download forkjoin-ai/mamba-2.8b-safetensors mamba-2.8b.knot --local-dir ./knots
```

Full format spec: [KNOT_FORMAT.md](https://github.com/forkjoin-ai/aether/blob/main/docs/model-cards/KNOT_FORMAT.md). Inspect the header with `bun run open-source/bitwise/scripts/dump-knot.ts mamba-2.8b.knot`.

## Deployment Architecture

This model runs on the **Aether distributed inference runtime** — a custom engine that shards model layers across multiple nodes for parallel execution:

1. **Coordinator** receives requests and manages token generation
2. **Layer nodes** each hold a subset of model layers (2 nodes for this model)
3. **Hidden states flow** between nodes via gRPC
4. **Zero cold start** via warm pool scheduling

Deployed via [Edgework.ai](https://edgework.ai) — bringing fast, cheap, and private inference as close to the user as possible.

## About

Published by [AFFECTIVELY](https://huggingface.co/forkjoin-ai) · Managed by [@buley](https://huggingface.co/buley)

We quantize and publish **production-ready models** for distributed edge inference via the [Aether](https://github.com/forkjoin-ai/aether) runtime. Every release is tested for correctness and stability before publication.

- [All models](https://huggingface.co/forkjoin-ai) · [GitHub](https://github.com/forkjoin-ai) · [Edgework.ai](https://edgework.ai)