Item detail

trotsky1997/OpenFugu

trotsky1997/OpenFugu is a open reverse-engineering of saka in RepoRadar's Apache-2.0 independent reimplementation of Sakan section, holding Gold tier and a 'try now' verdict. Its strongest signal is workflow potential, scored 9.1 out of 10.

Score8.0
Popularity183.0
Risklow
TierGold
Score breakdown
Usefulness8.0
Novelty9.0
Momentum7.0
Maturity8.4
Open-source/build8.4
Evidence7.2
Workflow potential9.1
Setup ease4.2

Popularity is tracked separately. Support, ads, sponsorships, and tips never affect these signals.

Why it matters

Useful for research groups and engineering teams who want to understand and replicate the Fugu mechanism (Sakana AI's policy-over-models orchestrator) without taking on a vendor lock-in: OpenFugu is the Apache-2.0 independent reimplementation from the Sakana papers + released artifacts, with the architecture, the training path, the serving endpoint, and the trained Conductor weights all on GitHub

Who should use it

Research groups and engineering teams who want to understand and replicate the Fugu mechanism (Sakana AI's policy-over-models orchestrator) without taking on a vendor lock-in: OpenFugu is the Apache-2.0 independent reimplementation from the Sakana papers + released artifacts, with the architecture, the training path, the serving endpoint, and the trained Conductor weights all on GitHub and Hugging FaceML researchers evaluating the per-query router pattern (the TRINITY coordinator is a ~0.6B backbone that never answers the user — it produces one hidden state at the penultimate token, a bias-free linear head scores each worker, the top worker is dispatched and its reply is returned; ~19.5K trainable numbers, the head + singular-value-fine-tuning offsets on 9 matrices, optimized gradient-free via sep-CMA-ES) against per-step coordination (the Fugu-Ultra Conductor emits a whole workflow DAG)Engineering teams who want an OpenAI-compatible orchestrator endpoint that takes one query and routes it to a litellm pool of workers (`openfugu/serve.py --slot-models <csv> --port 8088`) without writing the routing logic themselvesResearch groups who want to reproduce the measured per-question routing result (the trained router +107% over best single worker is published in `results/`) before betting a routing decision on the patternTeams who want to train a Conductor on their own dataset (the framework ships `train/train_conductor.py` with GRPO + the ToolScale dataset, and the published checkpoint on Hugging Face can be used as a starting point)Engineers who want to run the entire pipeline in one command (`python pipeline/e2e_train_serve.py` trains a fresh head, serves it, and verifies the result end-to-end) instead of chaining the four stages by handOrganizations that want a policy-over-models orchestrator they can audit end to end (the architecture is in `docs/HOW_FUGU_IS_IMPLEMENTED.md` with an EXEC/CODE/DATA evidence grade on every claim, and the weights are published so the routing behavior is reproducible)Engineering teams that want an Fugu-compatible adaptive k-of-n pool (`train/train_adaptive_pool.py` generalizes to arbitrary worker subsets, with the mock showing +44% over blind and 94% of oracle) for hot-swapping workers without retraining the headTeams evaluating Fugu-Ultra (the recursive topology where the Conductor revises its own output — `train/train_recursion.py` shows +9% over one-shot on a toy policy with headroom, and the held-out test in `eval/eval_recursion_real.py` shows TIE between round-0 and round-1, which is the honest negative result that the maintainer publishes) for test-time scaling

Who should skip it

Skip trotsky1997/OpenFugu for now if your priority is a tool you can use today without configuring a build pipeline or development environment.

About this signal

trotsky1997/OpenFugu is tracked by RepoRadar as a open reverse-engineering of saka in the Apache-2.0 independent reimplementation of Sakan section. It was first seen on 2026-06-25 and last updated on 2026-06-25. The current verdict is 'try now' with a Gold tier and hard setup difficulty. trotsky1997/OpenFugu leads on workflow potential (9.1) and novelty (9.0); its lowest signal is setup ease (4.2), so factor that in before investing setup time. This page summarizes the evidence RepoRadar has captured from captured source metadata. The score, tier, risk label, and verdict on this page are never influenced by sponsorship, ads, or tips — they reflect only the usefulness, popularity, novelty, momentum, maturity, and evidence signals described in the RepoRadar methodology.

How this item is evaluated

RepoRadar assigned trotsky1997/OpenFugu a composite score of 8.0 out of 10, placing it in the Gold tier. This score combines weighted sub-signals: usefulness (35%), novelty (18%), momentum (14%), maturity (10%), open-source/build quality (7%), evidence quality (6%), workflow potential (6%), and setup ease (4%). Popularity is tracked separately at 183.0 and never affects the composite score or tier. The risk label of 'low' reflects inherent user-impacting hazards, not generic novelty. Items with no risk flag may still require normal code review before production use.

Risk explanation

**Trained Conductor weights are under the Llama 3.2 Community License, not Apache-2.0.** The `NOTICE` is explicit: the OpenFugu code is Apache-2.0, but the trained Conductor weights published on Hugging Face (`huggingface.co/di-zhang-fdu/openfugu-conductor-3b`) are a fine-tune of Llama-3.2-3B-Instruct and carry the Llama 3.2 Community License. Adopters who want to ship a derived model need to comply with the Llama 3.2 license (acceptable use policy + the 700M-monthly-active-user clause for derivatives). Review the `NOTICE` license-layering section before redistributing any checkpoint; **Per-question routing +107% is query-level, not per-step coordination; Fugu-Ultra held-out shows TIE.** The eval is honest about scope: `eval/eval_orchestration.py` reports +107% over best single worker on the trained router, but the maintainer is explicit that this is query-level routing, not per-step coordination. The held-out Fugu-Ultra recursion test (`eval/eval_recursion_real.py`) shows TIE between round-0 and round-1, the honest negative result the maintainer publishes. Adopters betting a production routing decision on the pattern should reproduce the per-question orchestration eval on their own worker pool and read the Fugu-Ultra caveats before claiming the workflow-DAG pattern delivers the same lift; **Third-party material is fetched, not redistributed; verify the upstream license terms before distribution.** The OpenFugu code does not redistribute Qwen3-0.6B, `model_iter_60.npy`, or the 37-case fixture — `scripts/fetch_artifacts.py` pulls them from their licensed sources at install time. Adopters who want to redistribute the OpenFugu bundle need to verify the upstream license terms for each fetched artifact (Qwen3-0.6B's license, the original Sakana paper figures, the ToolScale dataset) and pass the licenses through to the redistribution.

Evidence links

Closest alternatives / related signals

openfugufugusakanasakana-aisakana-fugupolicy-over-modelsllm-orchestratororchestrator