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.
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.
$ 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 …
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.
Content hash, format inspection, tokenizer match and a six-prompt smoke test. A build that fails any gate never gets a page.
Each release ships the exact prompts, decoding parameters and judge model used for its model card. Reproduce a score or report a regression.
Checkpoints live in private object storage. Downloads require a license key, which keeps redistribution chains visible instead of anonymous.
A single Rust binary serves GGUF and safetensors on CPU, CUDA, Metal and Vulkan, with continuous batching, prefix caching and speculative decoding.
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.
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.
Four builds that cover the common cases: a general assistant, a code model, a retrieval stack and a frontier mixture-of-experts.
Balanced instruction model with reliable function calling and strict JSON adherence.
Repository-scale code model: multi-file context, tests first, minimal diffs.
Bilingual retrieval encoder, Matryoshka-trained to 256–2048 dimensions.
Frontier mixture-of-experts for agentic planning and structured extraction.
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.
# 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
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="")
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.
No key yet?
No editorial calendar. Notes appear when a measurement is worth sharing.
Perplexity moves by a fraction of a point after quantization. Production incidents do not.
Expert offload, page cache pressure and the routing behaviour that surprised us.
Why average latency stays comfortable right up until it does not.
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.
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.
68 contributors across inference, evaluation, release engineering and docs.
Access to weights on this mirror is granted by the node administrator after a manual review. Requests from anonymous sessions are queued and are not approved automatically.