Open-weight model hub

Download, inspect and serve open-weight models.

A small, independent catalogue of language, vision, code and embedding checkpoints. Every build is hashed, inspected and smoke-tested before publication, and every result in a model card can be reproduced with the harness we ship beside it.

15 checkpoints online GGUF & safetensors 3 licenses mirror node: this host
foundry — bash
$ foundry pull mf-ember-7b-instruct --quant q4_k_m
manifest   mf-ember-7b-instruct @ q4_k_m
sha256    9f4c…ae21  ok
tokenizer match  ok
smoke     6/6 prompts within tolerance  ok
auth      401 unauthorized

hint: this mirror serves weights only to licensed keys.
      sign in, then re-run the command.

$ foundry login --key $MF_KEY
$ foundry pull mf-ember-7b-instruct --quant q4_k_m
…
verified builds reproducible evals key-gated storage
15 checkpoints in the catalog
10,482,941 verified pulls served +8.4% this cycle
68 contributors
41 ms median TTFT, p95 fleet
What this mirror is

A catalogue you can audit, not a bucket you have to trust

The hub exists because model distribution is mostly unaccounted for. We publish the manifest, the digest, the evaluation harness and the license text for every checkpoint, and we keep the download surface boring.

Verified before published

Content hash, format inspection, tokenizer match and a six-prompt smoke test. A build that fails any gate never gets a page.

Numbers you can rerun

Each release ships the exact prompts, decoding parameters and judge model used for its model card. Reproduce a score or report a regression.

Key-gated weights

Checkpoints live in private object storage. Downloads require a license key, which keeps redistribution chains visible instead of anonymous.

One runtime, many shapes

A single Rust binary serves GGUF and safetensors on CPU, CUDA, Metal and Vulkan, with continuous batching, prefix caching and speculative decoding.

Bilingual by construction

English and Russian are sampled at parity in pre-training and reported separately in every evaluation, so non-English quality is measured rather than assumed.

Static mirror, boring ops

Catalog pages are static files behind a CDN edge. The mirror stays up when the inference fleet is busy, and availability is published on the status page.

Featured checkpoints

Most pulled this cycle

Four builds that cover the common cases: a general assistant, a code model, a retrieval stack and a frontier mixture-of-experts.

mf-ember-7b-instruct
Foundry Research · Chat model
MF-RL-1.0

Balanced instruction model with reliable function calling and strict JSON adherence.

Parameters
7B
Context
32k
Format
GGUF
Q4_K_M
4.4 GB
1,284,930 pulls Get weights
mf-cobalt-13b-code
Foundry Research · Code model
MF-CODE-1.0

Repository-scale code model: multi-file context, tests first, minimal diffs.

Parameters
13B
Context
64k
Format
GGUF
Q4_K_M
8.2 GB
459,912 pulls Get weights
mf-embed-1.2b
Foundry Research · Embeddings
Apache-2.0

Bilingual retrieval encoder, Matryoshka-trained to 256–2048 dimensions.

Parameters
1.2B
Context
8k
Format
GGUF
int8
0.7 GB
874,006 pulls Get weights
mf-opal-110b-moe
Foundry Research · Reasoning
MF-RL-1.0

Frontier mixture-of-experts for agentic planning and structured extraction.

Parameters
110B / 13B active
Context
256k
Format
GGUF
Q4_K_M
68.4 GB
96,340 pulls Get weights

All 15 checkpoints →

Quickstart

From catalog page to local inference in four commands

The CLI and the Python client speak the same manifest format. A license key unlocks the download; the runtime itself has no account requirement once the file is on disk.

  • Keys are per-workspace and can be rotated without re-downloading weights
  • Every pull verifies the manifest digest before writing to disk
  • The server exposes an OpenAI-compatible surface, so existing clients work unchanged

Full installation guide

terminal
# 1. install (MF_HOST = the host serving this page)
export MF_HOST="https://your-mirror-host"
curl -fsSL "$MF_HOST"/install.sh | sh

# 2. authenticate with a workspace key
foundry login --key "$MF_KEY"

# 3. pull a verified build
foundry pull mf-ember-7b-instruct --quant q4_k_m

# 4. serve it locally
foundry serve ./mf-ember-7b-instruct-q4_k_m.gguf \
  --ctx 32768 --parallel 4
quickstart.py
from modelfoundry import Client

client = Client()
client.login(api_key="MF-KEY-REDACTED")

for chunk in client.chat(
        model="mf-ember-7b-instruct",
        messages=[{"role": "user", "content": "hello"}]):
    print(chunk, end="")
Access model

Why the download asks you to sign in first

This mirror is not a public bucket. Checkpoints are stored privately and released against a license key tied to a workspace, which lets us rotate a compromised key without re-uploading a 68 GB checkpoint. Access requests are reviewed by the node administrator; the mirror does not grant keys automatically.

Free tiers cover research and internal evaluation. Hosted access and commercial redistribution are covered by the same key, with the terms recorded in the license file.

Sign in to the hub

No key yet?

  • Keys are revocable and scoped per workspace
  • Digest is checked on every pull, not just at publish time
  • No account is needed to read docs, cards or notes
Engineering notes

What the team is writing about

No editorial calendar. Notes appear when a measurement is worth sharing.

quantizationgguf

Quantizing models without losing the edge cases

Perplexity moves by a fraction of a point after quantization. Production incidents do not.

moeserving

Serving a 70B mixture-of-experts model on a single workstation

Expert offload, page cache pressure and the routing behaviour that surprised us.

performancelatency

Token economics: batching, the KV cache and the p99 cliff

Why average latency stays comfortable right up until it does not.

Release channel

What shipped recently

1.8.0 · current stable — catalog search, gated storage backend, Rust runtime 1.8.0.

1.7.2 · previous — tokenizer verification gate for converted checkpoints.

1.7.0 · previous — eval harness mf-eval/3.2 and per-language score reporting.

Full changelog

Community

Run by people who ship weights

A small collective of engineers and researchers. No growth team, no roadmap deck — the changelog is the roadmap, and the issue tracker is where model requests get argued about.

MEDKIVSVPRTB

68 contributors across inference, evaluation, release engineering and docs.