manifest-orchestrator-35b

<div align="center">

🚀 Manifest 35B

The frontier of the Manifest suite — enterprise-scale supply-chain intelligence

Built on Qwen3.5-35B-A3B · Mixture-of-Experts · ~36B total / ~3B active per token

     

</div>

---

Manifest 35B is MetaFloor's flagship supply-chain orchestrator — a general-purpose model that handles any supply-chain area: procurement, demand planning, warehouse operations, supplier relationship management, risk & resilience, transportation, inventory, and order fulfilment.

It is a LoRA fine-tune of Qwen3.5-35B-A3B — a large Mixture-of-Experts model (~36B total parameters, ~3B active per token) — trained on MetaFloor's full supply-chain instruction dataset.

Training

• Base model: Qwen3.5-35B-A3B (MoE, ~36B total / ~3B active)
• Method: LoRA fine-tuning (bf16), completion-only loss
• Adapter: rank 16 (~8.4M trainable params) on attention + MLP projections
• Data: 12,500 supply-chain instruction–response pairs spanning 8 sub-domains and every company scale
• Schedule: 3 epochs · final training loss ≈ 0.95
• Same reproducible pipeline used across the whole Manifest suite.

The Manifest family

Two kinds of models:

🧭 Orchestrators — general-purpose, handle any supply-chain area

🎯 Domain-experts — specialized for a single area

Orchestrators are scored on the general supply-chain benchmark; domain-experts on their focused domain benchmark.

Benchmark

Evaluated on 134 held-out supply-chain questions, scored blind by an independent LLM judge panel (glm-5.2 + gemma4 — a different model family), head-to-head against the base model:

Manifest 35B is preferred over the base even when the base is given a strong format prompt and few-shot examples — the panel's verdict: specialization confirmed, not just style.

Not just a model — a full stack

Manifest is backed by everything needed to build it and trust it:

• A purpose-built dataset — thousands of supply-chain instruction–response pairs spanning 8 sub-domains and every company scale, generated by a seed-driven operator-as-teacher pipeline.
• A reproducible training pipeline — documented LoRA fine-tuning.
• An independent benchmark — 116 held-out expert questions, scored blind by a panel of LLM judges.

We built the model, the data, and the evaluation.

How to use

Manifest 35B is a LoRA adapter over Qwen3.5-35B-A3B. Load the base model, then attach the adapter:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.5-35B-A3B"
model = AutoModelForCausalLM.from_pretrained(base, dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "metafloor-ai/manifest-orchestrator-35b")
tok = AutoTokenizer.from_pretrained("metafloor-ai/manifest-orchestrator-35b")

messages = [{"role": "user", "content": "How should we buffer inventory for a supplier with a 6-week lead time and 20% demand variance?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=512)[0][inputs.shape[1]:], skip_special_tokens=True))

The other orchestrators are also live:

• Manifest 0.8B — fast, low-cost, high-volume
• Manifest 2B — the balanced everyday default
• Manifest 4B — the most capable of the smaller tier

License

Manifest models and the supply-chain-eval benchmark are released under CC-BY-NC-4.0 — free for research and non-commercial use, with attribution. Commercial use requires a license from MetaFloor — get in touch at metafloor.ai.