bankai-v1 is a one-step agent orchestration model based on Qwen/Qwen3-Coder-Next, trained to perform one-step agent orchestration. Given a task, it selects one specialist worker and emits a complete delegated instruction as strict JSON.

This repository includes the original MLX LoRA, a llama.cpp GGUF LoRA, a fully fused 4-bit MLX model, and a fully fused Q4_K_M GGUF model. The adapter was trained against mlx-community/Qwen3-Coder-Next-4bit. The upstream model is an 80B MoE with approximately 3B active parameters.

Model Details

Property	Value
Model type	One-step agent orchestration model and adapters
Upstream architecture	Qwen3-Next hybrid Gated DeltaNet / attention MoE
Total base parameters	79,674,391,296 (79.674B)
Active parameters per token	Approximately 3B
LoRA trainable parameters	2,857,000 (2.857M)
Trainable share	Approximately 0.004% of base parameters
Base precision used for training	MLX 4-bit quantization
Adapter precision	Safetensors LoRA weights
Native base context	262,144 tokens
Fine-tuning sequence length	2,048 tokens
Languages	Primarily English; base model is multilingual
License	Apache-2.0
Framework	MLX LM 0.31.3

Qwen3-Coder-Next has 48 layers, 512 routed experts, ten activated routed experts per token, and one shared expert. bankai-v1 adapts only attention and gated-delta projections in the final 16 layers; it does not add LoRA matrices to the routed experts.

Repository Contents

Path	Type	Size	Standalone
`adapters.safetensors`	Original MLX LoRA	11,439,496 bytes	No
`adapter_config.json`	MLX LoRA configuration	1.3 KB	No
`gguf/bankai-v1-f16.gguf`	llama.cpp F16 LoRA	6,116,192 bytes	No
`gguf/bankai-v1-Q4_K_M.gguf`	Fused Q4_K_M GGUF	48,528,320,544 bytes	Yes
`mlx/`	Fused 4-bit MLX model, nine shards	Approximately 44.9 GB	Yes
`bankai-system.txt`	Exact training-time system prompt	1 KB	N/A
`training_config.yaml`	Reproducible training configuration	746 bytes	N/A
`evaluation.json`	Per-example behavioral evaluation	30 KB	N/A

Output Contract

json

{
  "action": "call_agent",
  "agent": "software_engineer",
  "model": "auto",
  "instruction": "A complete execution-ready instruction for the worker",
  "context_refs": ["user_request"],
  "expected_output": "completed_task_with_evidence",
  "budget": {"max_tokens": 4096}
}

Supported worker labels:

software_engineer
technical_writer
marketing_strategist
marketing_copywriter
sales_specialist
customer_support_specialist
data_analyst
product_designer
financial_analyst
compliance_specialist

Usage

Download the exact system prompt:

hf download ApplauseLab/bankai-v1 bankai-system.txt --local-dir bankai-v1

Fused MLX Model

The fused MLX model does not require an adapter:

bash

hf download ApplauseLab/bankai-v1 \
  --include "mlx/*" \
  --local-dir bankai-v1

mlx_lm.generate \
  --model bankai-v1/mlx \
  --system-prompt "$(cat bankai-v1/bankai-system.txt)" \
  --prompt "Design a GDPR-compliant CRM integration for a real-estate business." \
  --max-tokens 1024 \
  --temp 0

MLX Adapter

Install MLX LM on Apple Silicon:

bash

pip install "mlx-lm>=0.31.3" "huggingface-hub>=1.17"
hf download ApplauseLab/bankai-v1 \
  --include "adapter*" \
  --local-dir bankai-v1-adapter

Generate with the adapter:

bash

mlx_lm.generate \
  --model mlx-community/Qwen3-Coder-Next-4bit \
  --adapter-path bankai-v1-adapter \
  --system-prompt "$(cat bankai-system.txt)" \
  --prompt "Design and test a GDPR-compliant CRM integration for a real-estate business." \
  --max-tokens 1024 \
  --temp 0

For best consistency, use the included bankai-system.txt.

Fused GGUF / llama.cpp

The fused GGUF is standalone and does not require --lora:

bash

hf download ApplauseLab/bankai-v1 \
  gguf/bankai-v1-Q4_K_M.gguf \
  --local-dir bankai-v1

llama-cli \
  --model bankai-v1/gguf/bankai-v1-Q4_K_M.gguf \
  --system-prompt-file bankai-v1/bankai-system.txt \
  --prompt "Design a GDPR-compliant CRM integration for a real-estate business." \
  --temp 0 \
  --n-predict 1024 \
  --single-turn

GGUF Adapter

bankai-v1-f16.gguf is the same LoRA adapter converted for llama.cpp. It is not a standalone 80B model; load it alongside a compatible Qwen3-Coder-Next GGUF base:

bash

llama-cli \
  --model Qwen3-Coder-Next-Q4_K_M.gguf \
  --lora bankai-v1/gguf/bankai-v1-f16.gguf \
  --system-prompt-file bankai-v1/bankai-system.txt \
  --prompt "Design a GDPR-compliant CRM integration for a real-estate business." \
  --temp 0 \
  --n-predict 1024

The GGUF adapter was converted through PEFT tensor orientation with lora_alpha=128, preserving the MLX effective scale of 16 at rank 8.

Use the included bankai-system.txt for the exact worker vocabulary and JSON contract seen during fine-tuning. Shorter prompts may produce valid routing in a different schema.

Fusion Details

The fused MLX model was produced by merging the LoRA into mlx-community/Qwen3-Coder-Next-4bit and saving the resulting quantized MLX weights.

The standalone GGUF starts from the single-file Unsloth Q4_K_M conversion of Qwen3-Coder-Next. Untouched tensors are copied bit-for-bit. The 64 targeted attention and gated-delta matrices in layers 32-47 are dequantized to F32, the LoRA delta is applied at effective scale 16, and each matrix is requantized to its original Q4_K_M mixture type. The resulting GGUF contains 843 tensors.

Training

Training base: mlx-community/Qwen3-Coder-Next-4bit
Upstream base: Qwen/Qwen3-Coder-Next
Method: MLX QLoRA
Trainable parameters: 2.857M, approximately 0.004%
Adapted blocks: final 16 of 48
Adapted modules: full-attention Q/K/V/O projections and gated-delta input/output projections
LoRA rank: 8
LoRA scale: 16
Optimizer: AdamW
Learning rate: 1e-5
Updates: 612
Gradient accumulation: 2
Maximum sequence length: 2048
Training tokens: 355,178
Peak training memory: 74.4 GB on an Apple M4 Max

The source consisted of 1,080 synthetic task-to-delegated-prompt examples across ten task categories and twenty industries. Grouped splitting held out complete industries and complete (category, subtask) groups:

Train: 612
Validation: 224
Test: 244

Only assistant tokens contributed to the loss.

Evaluation

Final sampled validation loss: 1.606
Held-out test loss over 32 batches: 1.795
Held-out test perplexity: 6.018
Behavioral evaluation: one held-out example from each of ten categories
Strict JSON validity: 10/10
Runtime schema validity: 10/10
Worker-routing accuracy: 10/10
Complete contract: 10/10

The behavioral sample is small and measures formatting and category routing, not end-to-end quality of worker execution.

Both fused formats received inference smoke tests with the full system prompt. The fused MLX output matched the JSON contract. The fused GGUF output also matched the contract, selected compliance_specialist for a held-out real-estate GDPR/CRM/accounting request, and generated at approximately 64 tokens/second on an Apple M4 Max using Metal offload.

Limitations

This first version learns only single-step worker selection and prompt delegation. It does not execute workers, consume observations, choose among specific model providers, verify results, retry failures, call agents in parallel, optimize cost, or perform reinforcement-learned multi-turn orchestration.

Training examples are synthetic and stylistically concentrated. Outputs may be overly verbose and domain-specific legal, financial, medical, or compliance instructions require qualified human review.

Intended Use

Use bankai-v1 as a routing and prompt-delegation component in an agent runtime that validates its JSON, resolves the selected worker, enforces budgets, and records real worker observations. Suitable initial domains include software delivery, documentation, marketing, customer support, analysis, product design, finance, and compliance workflows.

Do not treat generated delegated instructions as professional legal, regulatory, medical, accounting, or investment advice. The adapter should not be given authority to execute high-impact actions without runtime controls and human review.

Training Data Notes

The source is a synthetic Cartesian task grid rather than organic production traces. Inputs combine 54 task types with 20 industries. Responses are execution-ready delegated prompts averaging roughly 345 words. A grouped split was used instead of a random split to reduce leakage from repeated industry and subtask templates. No private user conversations were used.

The source artifact has a documented lineage discrepancy: records task-0881 through task-0900 differ from their current shard copies. Training used the checksummed final training_set.jsonl artifact.

Integrity

File	SHA-256
`adapters.safetensors`	`66df3ea1d9a9f0b1c3e07b1620a33dfb0b3bf36b6ea4b1e52fe2fc1bd8da67d9`
`gguf/bankai-v1-f16.gguf`	`184b7855166c6c97fe366b4412de51cf98474dfc5df7caa95bda09f83c80e84d`
`gguf/bankai-v1-Q4_K_M.gguf`	`75f8200b83756e6f32565b4579c9bb07aa6aa591fca380bc3f85af1138102296`

The fused GGUF also passed llama.cpp per-tensor hashing with whole-file XXH64 c64819733a364e4d.