OpenTransformer/AGILLM-4.3

AGILLM 4.3

AGILLM 4.3 is the AGILLM 4.2 warm start with shared MoE experts and DiffusionBlocks training. The compatibility runtime file is still named agillm41.py, but this repo tracks the 4.3 runtime and public volunteer path.

Public Safety Boundary

This public repo is for inspection, local inference experiments, and untrusted volunteer helpers. It intentionally excludes trusted-core operations, private topology, watchdog launch scripts, live hot configs, SSH paths, API tokens, and checkpoint merge scripts.

Untrusted volunteer nodes should use only the outbound public join flow. The published coordinator is https://join.opentransformers.online, and its health endpoint is https://join.opentransformers.online/health.

python public_join/agillm41_join_worker.py \
  --coordinator-url https://join.opentransformers.online \
  --device cpu \
  --threads 2 \
  --loop

The worker opens outbound HTTPS only, verifies SHA-256 for lease artifacts, receives short-lived lease tokens only, and submits results to quarantine. Public helper results must be validated before they can affect a checkpoint.

Files

• agillm41.py: latest public AGILLM runtime, including AR/SAT/NAT inference and DiffusionBlocks paths.
• public_join/: outbound worker, public lease coordinator, and public validation/points helpers.
• agillm4/training_bench/agillm4_slice_bench_worker.py: default public slice worker for leased training packages.
• distributed_infer/: public distributed inference harnesses without private launch topology.

Private Counterpart

Trusted-core operations live in the private repo Marxist-Leninist/agillm4.3-private and private HF repo OpenTransformer/agillm4.3-private.

Hugging Face

Public model card and checkpoint lineage: https://hf.co/OpenTransformer/AGILLM-4.3

Run Your Own Federated Network

If you want to host your own decentralized training swarm for AGILLM models, you can run the volunteer coordinator and validation endpoints yourself.

1. Start the Network Host (Coordinator): This script runs a FastHTML web server that distributes training leases to workers and receives asynchronous .pt gradient updates.

   python public_join/agillm41_network_host.py \
     --host 0.0.0.0 \
     --port 8787 \
     --spool ./agillm41_lease_spool \
     --public-base-url http://YOUR_IP:8787
   

2. Add Packages (Master Node): Your master training loop exports bench packages which are added to the spool:

   python public_join/agillm41_network_host.py add-lease \
     --spool ./agillm41_lease_spool \
     --package /path/to/exported_bench_pkg.pt \
     --base-ckpt /path/to/base_model.pt
   

3. Validate Results: Once workers submit results to the quarantine directory in the spool, you must validate them before your master applies them.

   python public_join/agillm41_validate_and_credit.py \
     --spool ./agillm41_lease_spool \
     --base-ckpt /path/to/base_model.pt
   

   Validated updates will be moved to an accepted/ directory which your master can then asynchronously merge.

Latest Inference Smoke Test

This is a raw quality smoke test from the active AGILLM-4.3 training checkpoint. It is included to show that the checkpoint loads and decodes; the text is not yet meant to be good instruction-following output.

• Created UTC: 2026-06-18T05:13:28Z
• Checkpoint: final.pt
• Checkpoint step: 2161945
• Method: ar_block_stream_fp16_sdpa_ignore_eos
• Decode stats: 55.85s | 64 tokens | 1.1 tok/s
• Full artifact: inference_samples/step2161945_20260618T051328Z/README.md

Prompt

In one short paragraph, explain what AGILLM4.3 is doing during this training run.

Output

If you can be a lot of the same time to do that we are not like it's I'm so if they have some things and you don't know about your way for my book.

I think how you want to get me! You're just in our thing, I've never had been on the other