Hugging Face index · DeltaVerse

Our work on Hugging Face (Spaces, datasets and models), and how mindX ingests from inference, optimizing for all advantages. Statuses were read from the Hub at 2026-10-12 23:19 UTC. Machines can read the same index as hfindex.json.

Spaces

Every page is built on the ultimate input field: response windows you can move, resize and pop out, docking to any edge, T (terminal) mode, and .history / .memory kept in bankML's formats.

Gregory-L/Savante

Savante, interactive edition

Savante's page with the ultimate input field: sAGI mode (Savante), a generic mode and T (terminal) mode. It has four carriers: bankML in the browser (browseML, with receipts), Bonsai on the free public CPU, Bonsai on your own machine, and a Hugging Face provider. Delivery records are kept in .memory, and the improve.skill count is also shown in 18-decimal EVM units.

cost
free (browseML, free CPU); provider: visitor's credits
status
running · static
commit
74d71c3 · 2026-10-09

PYTHAI/savante

Savante, the template

The public edition and template of Savante's office: the rooms Ask, Hub, Office, Verdict, Integrity and Skills, with Bonsai on a free CPU. Kept exactly as it is.

cost
free; provider: visitor's credits
status
running · static
commit
66625db · 2026-09-26

Gregory-L/ultimate-webgpu

ultimate WebGPU

The ultimate input field on the visitor's GPU: Qwen3-0.6B text and Agate images (Logolabs, MIT, marked AI-generated) on WebGPU, with bankML as a separate, verified CPU engine (browseML, with receipts).

cost
free (runs in the visitor's browser)
status
running · static
commit
65bce6f · 2026-10-09

Gregory-L/zeroscopeprivatenot read now

zeroscope: text to video

Text to video in the ultimate input field. Wan 2.2, Wan 2.1, HunyuanVideo and LongCat-Video run at a Hugging Face Inference Provider, and each clip opens in a response window with a save link.

cost
visitor's Hugging Face credits
status
running · static
commit
6a52cba · 2026-10-09
read
last seen 2026-10-09 18:51 UTC; the Hub answered HTTP 401: private or absent, and no token was given

Gregory-L/platypusaiprivatenot read now

platypus

Mistral-7B-v0.1-Open-Platypus as a Q4_K_M GGUF pinned by sha256, served by llama.cpp b11515 on the free CPU behind the ultimate input field. It is bankML-ready: bankML takes over once it has SentencePiece and Q4_K.

cost
free (the Space's CPU)
status
running · cpu-basic · waiting for its model conversion
commit
27dad1f · 2026-10-09
read
last seen 2026-10-09 18:51 UTC; the Hub answered HTTP 401: private or absent, and no token was given

Gregory-L/mindXhfgradio

mindXhfgradio

mindX's Gradio Space: the mindX coach as one UI, on ZeroGPU hardware. The rooms are Door, mindX, Workbench (imprinted generation against its untouched SmolLM2-135M base, influence = after − before), Table, Scales, Kiln · mindXtrain, Map · Cabinet, Ledger and ternary · speak. Its /mindx endpoint runs Bonsai-8B (1-bit) and Ternary-Bonsai-8B through llama.cpp b11192 on the Space's CPU, both pinned by sha256: that is the free public carrier for Savante and Jaimla, whose sAGI and jaimla answers are recorded to their public loops. The `here` backend runs a mindXascension generation in-process on ZeroGPU on the signed-in visitor's quota. MCP server: probe and plan are tools.

cost
free (Bonsai on the Space's CPU); `here`: the visitor's ZeroGPU minutes
status
running · zero-a10g
commit
b0cd686 · 2026-09-27

PYTHAI/bankml

bankML

bankML: verified 1-bit and ternary LLMs on a CPU, in Rust, bit-exact against llama.cpp b11192. The page asks your own bankml serve (a receipt whose sha256 the page checks) or, signed in, a Hugging Face provider with bankML's persona, labelled as not bankML. It holds bankML's whole source and the Space mount (uif-space.js) the ultimate-input-field pages use.

cost
free (your own bankml serve); provider: visitor's credits
status
running · static
commit
36a3c9b · 2026-10-10

PYTHAI/bankML-UIF

bankML UIF

bankML behind the ultimate input field, as a React app: a draggable, resizable, dockable field with a chat mode and T (terminal) mode, and output fields you can spawn and pop out. A first visit answers with bankML in the browser (browseML, Bonsai-1.7B Q1_0 checked against its sha256 pin, a receipt on every answer); your own bankml serve and a Hugging Face provider are the other two answerers. Formerly PYTHAI/ultimate-bankml-ui (the npm package keeps that name).

cost
free (browseML, your CPU); provider: visitor's credits
status
running · static
commit
d3ab583 · 2026-10-10

PYTHAI/mindX

mindX, the front door

mindX's static front door on Hugging Face, in the first person: a self-building cognitive architecture whose BDI agents do the work, audited against eight Gödel-machine predicates, dreaming over its own memory and training its own weights on a two-core CPU. Every number is read live from mindx.pythai.net; it also renders the Thesis, Manifesto and Navigation from PYTHAI/mindX-docs, hashed in the browser.

cost
free
status
running · static
commit
77a7af5 · 2026-10-10

PYTHAI/browseML

browseML

bankML's own engine compiled to WebAssembly, on the visitor's CPU: Bonsai-1.7B Q1_0 (248 MB, cached once) checked against its sha256 pin before it answers, and a receipt on every answer that the page recomputes. No server answers; the Space only serves files.

cost
free (runs in the visitor's browser)
status
running · static
commit
1f9a158 · 2026-10-10

PYTHAI/minaiml

minaiml

minaiml (say it minimal), PYTHAI's delivery layer: software that runs reasonable-sized models, mixture-of-experts first, on laptops and phones people already own. Labelled LAB TEST · ALPHA: it has never generated a token on a phone, and every figure on the page is a model or someone else's published measurement.

cost
free
status
running · static
commit
5f02c9e · 2026-09-22

PYTHAI/README

PYTHAI organization card

The PYTHAI organization card (static): what PYTHAI is, start-here links to mindX, bankML and Savante, and a deeper full.html. Its "everything else" row links to this index.

cost
free
status
running · static
commit
5181efa · 2026-10-09

Gregory-L/mindX-ascend

mindX-ascend (placeholder)

Still the Hugging Face static template (Transformers.js object detection with Xenova/detr-resnet-50); no mindX content. The gen-39 weights it was named for are in the dataset Gregory-L/mindX-ascend-weights.

cost
free
status
running · static
commit
af0233e · 2026-09-02

Gregory-L/ZENMLOS

Zenml Server (2023)

A 2023 duplicate of zenml/zenml (Docker). Not part of mindX; sleeping.

cost
free while asleep
status
sleeping · cpu-basic
commit
bbabff9 · 2023-06-02

Datasets and models

PYTHAI/savante-loop

savante-loop

Savante's improve.skill loop: .memory, .history and publication.json, plus agents/: ERC-8004 registration cards for the sAGI family (savante, jaimla, luvai, codephreak, simplecoder, bankml) with agents/index.json. None is minted.

cost
free
status
published
commit
0f755f6 · 2026-10-12

PYTHAI/jaimla-loop

jaimla-loop

Jaimla's improve.skill loop, which runs fully automatically: .memory, .history, SKILL.md, the archived skill versions, operator.log and an audited dataset card.

cost
free
status
published
commit
6d28897 · 2026-10-09

PYTHAI/Bonsai-8B-gguf-fork

Bonsai-8B-gguf-fork

Bonsai-8B Q1_0 (1-bit), pinned by sha256: Savante's and Jaimla's carrier.

cost
free
status
published
commit
87c4ff9 · 2026-09-26

Gregory-L/Mistral-7B-v0.1-Open-Platypus-GGUFnot read now

Mistral-7B-v0.1-Open-Platypus-GGUF

Platypus Q4_K_M with FORK.json, published by Gregory-L/platypusai on its first run.

cost
free
status
not yet published
commit
—
read
last seen 2026-10-09 18:51 UTC; the Hub answered HTTP 401: private or absent, and no token was given

PYTHAI/mindXascension

mindXascension, the lineage

mindX's training lineage in the open: weights/genN/ (merged model, adapter and provenance, generations 1–41) and machine.dream/, the curated corpus of mindX's own memory in its own voice. Base HuggingFaceTB/SmolLM2-135M, Apache-2.0. The gate refuses more than it accepts: generation 39 is the newest it accepted.

cost
free
status
published
commit
4bd31b9 · 2026-09-26

PYTHAI/mindXascensionprivatenot read now

mindXascension, the adapters (private)

The private model repo that receives every generation's adapter (genN/). It holds generations the public dataset does not, so a token-less reader can load only up to generation 41.

cost
owner only
status
private · gen 1–74 as read 2026-09-10
commit
—
read
last seen CLAUDE.md, 2026-09-10; the Hub answered HTTP 401: private or absent, and no token was given

PYTHAI/mindX-docs

mindX-docs

mindX's documentation mapped for mindXtrain: the Thesis, Manifesto and Navigation with a mapping, each linking back to its page on mindx.pythai.net and to the read-aloud on the node.

cost
free
status
published
commit
247f2f3 · 2026-09-16

PYTHAI/bankonos-images

bankonOS images

OCI archives built by bankonOS's podman/deployer (alpine-base, debian-base, bitcoind, ipfs, openbdk), each beside a sha256 and a manifest.json naming its template, variables and bankonOS commit. No image carries a secret.

cost
free
status
published
commit
07cb54e · 2026-09-26

Gregory-L/machine.dream

machine.dream (early)

An early (2026-09-02) upload of machine.dream training JSONL from mindX's agents. The curated corpus now lives in PYTHAI/mindXascension/machine.dream. No dataset card.

cost
free
status
published
commit
2677f2b · 2026-09-02

Gregory-L/mindX-ascend-weights

mindX-ascend-weights (gen 39)

Generation 39 only: the LoRA checkpoint and the Ollama merge. Superseded by PYTHAI/mindXtrain39 and PYTHAI/mindXascension. No dataset card.

cost
free
status
published
commit
dc76d92 · 2026-09-02

PYTHAI/mindXtrain39

mindXtrain39

Generation 39 of the dream→weights lineage, the newest the proof-of-recall gate accepted (imprint Δ recall +0.1002): SmolLM2-135M with a LoRA r=16 merged, trained on two vCPUs with no GPU. Merged weights, adapter/, training log, a THOT and iNFT facets. No inference provider serves it.

cost
free
status
published
commit
0ea7d6b · 2026-09-12

PYTHAI/mindXtrain

mindXtrain

mindXtrain for mindX, the Hugging Face fork (from the GitHub upstream at 661bd41, which is now archived). New work on the training framework lands here. Code, not weights.

cost
free
status
published
commit
9c5dc6c · 2026-10-05

PYTHAI/SAVANTE

SAVANTE, the .persona

Savante (savante_sagi, the prototype sAGI) as a .persona: no weights, and never will be. The charter, facet bundle, integrity ledger and proofs, mirrored from github.com/cryptoAGI/savante under canon/. mindXhfgradio reads Savante's system prompt from here.

cost
free
status
published
commit
df740f7 · 2026-09-26

PYTHAI/voaice

voaice, the voice library

voaice's voice library: 70 open-licensed Piper voices in 29 language families, unchanged, with their licences, and cards/ with 81 voice cards (70 Piper + 11 in-house), each stating its licence and whether it is mintable. Summarised under Voices below.

cost
free
status
published
commit
5664171 · 2026-10-08

PYTHAI/chronos-oracle-llm-tools

chronos.oracle LLM tools

The tools mindX measures itself with: time to first token, generation, wall time, tokens, CPU and GPU time per LLM call, every event stamped with chronos.oracle time (kairos_meter.py and friends). Code and measurements, no weights.

cost
free
status
published
commit
065e8ee · 2026-09-26

Collection: platypus (akjindal53244/Mistral-7B-v0.1-Open-Platypus), served by Gregory-L/platypusai.

Licence-locked pointer forks

Each preserves an upstream model's licence, configuration, tokenizer and code at one exact commit, with SHA-256 digests. The weights stay at the source: these are pointers, not copies.

Voices

The open-voice collection in PYTHAI/voaice: 70 Piper voices in 29 language families and 11 in-house voices, each with a card stating its licence and whether it is mintable (read now from cards/index.json). Mintable means the licence would allow it; nothing here is minted.

Piper voices by language

70 voices · 70 of 70 mintable

  • English 11
  • Dutch 7
  • Spanish 5
  • French 5
  • Ukrainian 5
  • German 4
  • Kazakh 3
  • Catalan 2
  • Greek 2
  • Basque 2
  • Persian 2
  • Italian 2
  • Swedish 2
  • Telugu 2
  • Urdu 2
  • Bengali 1
  • Czech 1
  • Estonian 1
  • Finnish 1
  • Armenian 1
  • Lithuanian 1
  • Latvian 1
  • Marathi 1
  • Nepali 1
  • Norwegian 1
  • Polish 1
  • Portuguese 1
  • Slovenian 1
  • Chinese 1

Piper voices by licence

every one an open licence; credit Piper and each voice's dataset

  • CC BY 4.0 International 26
  • CC0 1.0 Universal (public domain dedication) 19
  • Public domain 9
  • Apache License 2.0 5
  • CC BY-SA 4.0 International 4
  • CC BY-SA 3.0 Spain 2
  • MIT License 2
  • The Unlicense 1
  • GNU AGPL v3 1
  • GNU GPL v2 1

In-house voices

11 cards · 5 mintable

The espeak-ng identities are voaice's own measurements and are mintable. The voices rendered from lessac-derived Piper bases are kept as the realm's expressions and are not mintable; vclone is a template with nothing measured yet.

  • voaice — JAIMLA (espeak-ng) jaimla-espeak mintable
  • voaice — NEURAL (espeak-ng) neural-espeak mintable
  • voaice — OVERLORD (espeak-ng) overlord-espeak mintable
  • voaice — OVIE (espeak-ng) ovie-espeak mintable
  • voaice — PARTICIPANT (espeak-ng) participant-espeak mintable
  • voaice — JAIMLA jaimla not mintable
  • voaice — LEADER leader_tone not mintable
  • voaice — LEADER leaderofearth not mintable
  • voaice — NEURAL neural not mintable
  • voaice — OVERLORD overlord not mintable
  • voaice — vCLONE vclone not mintable

SAVANTE's gene pool

SAVANTE's voice is being evolved by survival of the fittest from this pool: recipes are bred from the lineage, each render is scored by measured fitness (intelligibility by whisper.cpp, presence and timbre against the target, CPU cost), and the fittest is kept. NEURAL and JAIMLA are the two fixed references, set in stone; the target lies between them.

The references, set in stone

templates from the personal collection, never retuned

NEURAL
en_GB-alan-medium · 98 Hz · the male reference — the voice that does not move
JAIMLA
en_GB-jenny_dioco-medium · 182 Hz · the female reference — the first voaice, whom SAVANTE is from
target
156 Hz, 0.75 of the way from NEURAL to JAIMLA · centroid 2748 Hz
fitness
intelligibility 0.35 · presence 0.2 · timbre 0.1 · efficiency 0.15 · fidelity 0.2

The open lineage mintable

11 open-licence English voices

Built only from open-licence voices in PYTHAI/voaice. A SAVANTE bred from this lineage is mintable; attribution follows each layer's licence.

  • en_GB-cori-high · public domain
  • en_GB-cori-medium · public domain
  • en_GB-southern_english_female-low · CC-BY-SA 4.0 International
  • en_US-bryce-medium · public domain
  • en_US-john-medium · public domain
  • en_US-kathleen-low · CC0
  • en_US-kristin-medium · public domain
  • en_US-libritts-high · CC BY 4.0 · not bred from yet
  • en_US-ljspeech-high · public domain
  • en_US-ljspeech-medium · public domain
  • en_US-norman-medium · public domain

The in-house lineage not mintable

an option, kept as the realm's expressions

SAVANTE extrapolated entirely from the personal collection: NEURAL and JAIMLA as templates, LEADER's broadcast desk and OVERLORD's delivery lock as technique. Lessac-derived bases, so not mintable.

  • en_GB-jenny_dioco-medium as JAIMLA
  • en_GB-alan-medium as NEURAL
  • en_GB-northern_english_male-medium as sAGI

How mindX ingests from inference: optimizing for all advantages

Every route has an advantage of its own: a receipt, zero cost, privacy on the visitor's device, speed and scale, or history. mindX weighs them together and takes the route that serves the question best, never one advantage at the expense of the rest.

A loop mindX can run
  1. Read hfindex.json and weigh each route by its advantage, cost and auth.
  2. Choose by advantage: bankML on the node when the answer must be verified (a receipt), mindXhfgradio or Platypus for public answers at no cost, the browser engines when the exchange should stay on the visitor's device, Inference Providers for speed and large models when the token's owner has agreed to pay, and the loops' datasets for history.
  3. Ask, and keep every answer as a .memory record: question, answer, carrier, token counts, timings and response_sha256, the same fields Savante's improve.skill reads.
  4. Check the record: a receipt where the route gives one (bankML), the engine's token counts where it doesn't, and finish == "stop". Anything else counts as a draft, never a finding.

The loops' data (no inference)

advantage
history: every recorded answer, with no inference at all
cost
free
auth
none (public datasets)
endpoint
https://huggingface.co/datasets/PYTHAI/savante-loop/resolve/main/{.memory,.history,publication.json,agents/index.json}

Download the JSONL files and read them line by line. .memory is the record of each interaction (question, answer, carrier, token counts, timings, response_sha256), .history is derived from it, and agents/index.json lists the agent cards with their CIDs.

from huggingface_hub import hf_hub_download
import json
p = hf_hub_download("PYTHAI/savante-loop", ".memory", repo_type="dataset")
records = [json.loads(l) for l in open(p) if l.strip()]

Bonsai on mindXhfgradio (free public CPU)

advantage
no cost and no auth: public Bonsai answers that feed the improve.skill loop
cost
free
auth
none (public Space)
endpoint
https://gregory-l-mindxhfgradio.hf.space · Gradio API /mindx

Call the Space with gradio_client. Set `bz_persona` to "sAGI · Savante on Bonsai", or to "the system box above" with your own `system` prompt. `where` chooses bonsai-8b or ternary-bonsai-8b. Answers come one at a time on a shared CPU (about 5–20 s warm, minutes when the Space is cold). It's a public Space, so send nothing private: sAGI answers there feed the public improve.skill loop.

# pip install "gradio_client>=1.4"   (checked with 2.7.2; 0.x cannot read this Space's API)
from gradio_client import Client
c = Client("Gregory-L/mindXhfgradio")
text, stats_html, _ = c.predict(text="What is sAGI?", where="ternary-bonsai-8b (CPU llama.cpp)",
    bz_persona="sAGI · Savante on Bonsai", npred=256, temp=0.3, top_p=0.9, rg_use=False, api_name="/mindx")
# generic: bz_persona="the system box above", system="You are a helpful, concise assistant."
# stats_html: the engine's own counts (prompt · completion · total, tok/s, llama.cpp build)

Platypus on Gregory-L/platypusai (free CPU, llama.cpp)

advantage
no cost: a 7B instruction model behind an OpenAI-compatible endpoint
cost
free
auth
a Hugging Face read token (the Space is private)
endpoint
https://gregory-l-platypusai.hf.space/v1/chat/completions (OpenAI-compatible), /completion, /health

Once the model is converted (status.json: state ready), llama-server answers on the Space's CPU at about 2–4 tokens a second. Use the Open-Platypus format (### Instruction / ### Response) with /completion, or the chat endpoint. Calling a Space is not billed as inference.

curl -s https://gregory-l-platypusai.hf.space/completion \
  -H "Authorization: Bearer $HF_READ_TOKEN" -H "Content-Type: application/json" \
  -d '{"prompt":"### Instruction:\n\nWhat is 17 * 23?\n\n### Response:\n","n_predict":128,"stop":["### Instruction:"]}'

bankML or llama.cpp on mindX's own node

advantage
verification and privacy: a receipt on every answer, a model pinned by sha256, nothing leaves the node
cost
free (your hardware)
auth
local, or an SSH tunnel
endpoint
bankml serve (http://127.0.0.1:18093) · llama-server (http://127.0.0.1:18092)

The verified route. `bankml serve` puts a receipt (response_sha256) on every answer and pins its model by sha256. Ingest the answer, check the receipt and keep the record in .memory.

curl -s http://127.0.0.1:18093/v1/chat/completions -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"who are you?"}],"max_tokens":128}' | jq '.bankml_receipt'

The browser engines (browseML and WebGPU)

advantage
privacy and no server cost: the model runs on the visitor's own device
cost
free (the visitor's device)
auth
none
endpoint
Gregory-L/Savante · Gregory-L/ultimate-webgpu: they run in the visitor's browser

These engines can't be called from a server: they answer only in the browser that loaded them. mindX ingests what they produce instead. The ultimate input field keeps every exchange in .history and every delivery in .memory (in the same shape as the loops' .memory, never sent to a model), and both export as JSONL from the field's Windows menu.

# the field's Windows ▸ .memory ▸ export → memory.jsonl
records = [json.loads(l) for l in open('memory.jsonl') if l.strip()]
deliveries = [r for r in records if (r.get('source') or {}).get('kind') == 'delivery']

Hugging Face Inference Providers

advantage
speed and scale: fast, large models on demand
cost
a small monthly free allowance, then paid
auth
the caller's token, billed to the caller
endpoint
https://router.huggingface.co/v1/chat/completions

Its advantage is speed and model size. It carries a small monthly free allowance and then bills the caller, so mindX takes this route when speed or scale is worth the cost and the token's owner has agreed to pay. Gregory-L/zeroscope and the provider carrier on Savante bill the visitor who signs in.

# billed to the caller's token beyond the monthly allowance
curl -s https://router.huggingface.co/v1/chat/completions -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"<model id>","messages":[{"role":"user","content":"What is sAGI?"}],"max_tokens":256}'