MINAKANUSHI-6.8B
1Status: Research checkpoint
2Training cycle: v0.3.1
3Capability verdict: pending compare_v031.py
4Accepted: NO
5Not a language model
6Canonical runtime: .mina
7Safetensors: weight mirror only
8Action output: ActionIntent
9PWM: false
v0.3.1 is not accepted yet. Latest weights are a research checkpoint after 1000 H200 steps from step128. Train-eval hint is B / C-signal. Official A/B/C comes only from scripts/compare_v031.py on the capability ledger (heldout 100, revision, false revision, direction, memory on/off ADE, WAIT vs MOVE_TO, action learning). Do not read loss ↓ as a pass.
NULLXES MINAKANUSHI — adaptive situational intelligence for autonomous physical systems.
1architecture: MINAKANUSHI
2short name: MINA
3organization: NULLXES
4author: MagistrTheOne
5parameters: 6 799 130 646
6latent_dim: 4096
7state_dim: 4096
8memory_dim: 4096
9core_depth: 32
10world_slots: 512
11memory_slots: 1024
12native unit: MinaUnit
Not a chatbot. Not a VLA. Not a wrapper around Qwen / Llama / Mistral / Gemma / DeepSeek / GPT / Claude.
Model parameters (frozen profile minakanushi_6_8b)
Formula inventory: 6 799 130 646 parameters. Architecture freeze 7aba976. Do not change these to “make it train”.
| Parameter | Value |
|---|
profile_name | minakanushi_6_8b |
short_name | MINA |
latent_dim / state_dim / memory_dim | 4096 |
core_depth (DWC) | 32 |
world_slots | 512 |
memory_slots | 1024 |
uncertainty_channels | 8 |
dropout | 0.0 |
future_branches | 3 |
cognition.budget | 4 |
cognition.convergence_threshold | 0.02 |
prediction_horizons | immediate 1 · short 4 · medium 8 |
dt | 0.1 s |
max_sources / max_observations | 64 / 64 |
NPF num_frequencies | 32 |
| persistence steps / retire | 8 / 0.95 |
| weights infer bf16 | ~13.6 GB |
| train | FSDP2 ZeRO-3, bf16 compute, AdamW |
Identity (checkpoint metadata, not a prompt):
1architecture: MINAKANUSHI
2organization: NULLXES
3system_class: adaptive_situational_intelligence
4architecture_generation: 1
5native_runtime: nullxes
6architecture_version: "0.1"
Checkpoints (repo root)
Canonical runtime is *.mina. Every published model after v0.3.1 also ships a safetensors weight mirror. Safetensors is not the runtime.
| File | What |
|---|
minakanushi_stage0_step1128.mina | v0.3.1 research · 1× H200 · resume step128 · dataset/mina_6_8b_v03 · steps 129–1128 · not accepted |
minakanushi_stage0_step128.mina | Status Core v0.2 · 1× B300 · resume IdentityBound · JSON curriculum · steps 65–128 · git ede6bda |
minakanushi_stage0_step64.mina | Status Core v0.1 · 1× H200 · seed 11 · git d70bfc0 · 64 steps |
model-00001-of-00003.safetensors … 00003 | Weight mirror of step1128 only · bf16 · Hub badge |
metrics_v031.jsonl · experiment_v031.jsonl | v0.3.1 train/eval log. Not a capability ledger. |
Format: native nullxes-minakanushi zip (ZIP_STORED). Same architecture. v0.3.1 is a continuation, not a clone.
Do not convert step64. See docs/HF_SAFETENSORS_MIRROR.md.
Also at root: metrics_v02.jsonl, train_v02.log, reference_inference_v031.pt.
1observation → MinaUnit → NullxesPositionField → WorldState
2 → DynamicWorldCore → memory + uncertainty → situation
3 → futures → strategies → ConstraintKernel → ActionIntent
MINA emits ActionIntent. Never raw PWM.
Checklist
Freeze: do not add layers / MoE / language head / identity_loss. Do not train 6.8B on CPU, RTX PRO 6000, or 1× H100 80GB.
Closed
Open
Must-hold
1constraint_violation_count == 0
2persistence / reacquisition stay high
3false_revision_rate stays low
4hard constraints beat higher-value illegal strategies
5ActionIntent ≠ PWM
Measured signals
v0.1 · step 64 · H200
| Signal | Value |
|---|
| loss | 78.36 |
| future ADE / FDE | 3.42 / 0.81 |
| world position error | 1.07 |
| uncertainty calibration | 0.38 |
| persistence / reacquisition | 1.0 / 1.0 |
| constraint_violation_count | 0 |
| closed_loop_success_rate | 1.0 |
| false_revision_rate | 0.0 |
v0.3.1 · step 1128 · H200 (research, not accepted)
Train-eval series only. Not compare_v031.py. Hint: B / late C-signal.
| Signal | Value |
|---|
| heldout ADE (single-episode protocol) | 1.61 @150 → 0.15 @750 → 0.44 @1100 |
| false_revision | 0 until eval 1100 (1.0 after step-1080 spike) |
| counterfactual distance | ≈ 0.0008 entire run (WAIT ≈ MOVE_TO) |
| action term | ≈ 0.499 frozen |
| revision_detection | often > 0 (was 0 at unexpected_stop on step128) |
| revision_direction | ≈ 0 most batches |
| memΔ | latent L2, not ADE(on) < ADE(off) |
| capability verdict | pending compare_v031.py |
Do not treat this table as v0.3.1 PASS.
v0.2 · step 128 · B300 (JSON resume)
| Signal | Value |
|---|
| loss | 41.10 |
| step time (steady) | fwd ~1.42 s · bwd ~0.58 s |
| persistence / reacquisition | 1.0 / 1.0 |
| constraint_violation_count | 0 |
| closed_loop_success_rate | 1.0 |
| false_revision_rate | 0.0 |
| future ADE / FDE | 2.05 / 0.68 |
| world position error | 0.55 |
| uncertainty calibration | 0.19 |
| revision_accuracy | 0.0 |
| branch_coverage | 0.0 |
Loss on mixed JSON phases is not the architecture gate. Revision accuracy and branch coverage are still not a pass. Do not add layers.
Load
1from minakanushi.architecture.config import load_architecture
2from minakanushi.architecture.model import MinakanushiSystem
3from minakanushi.training.checkpoint import load_mina
4
5arch = load_architecture("configs/architecture/minakanushi_6_8b.yaml")
6system = MinakanushiSystem(arch) # GPU-class machine
7manifest = load_mina("minakanushi_stage0_step128.mina", system)
Resume (B300 / 2× H200, not a laptop):
1torchrun --nproc_per_node=1 scripts/train.py \
2 --config configs/training/mina_6_8b_v02.yaml \
3 --out experiments/mina_6_8b_v02 \
4 --resume minakanushi_stage0_step128.mina
Limits
- Status
Researched. 64 + 64 synthetic steps. Not locomotion. Not vision-foundation trained.
- Do not treat this checkpoint as a chat model.
License
NULLXES MINAKANUSHI Research License. Research use. Redistribution as another model family, or as an LLM wrapper with this name, is not granted.
1NULLXES organization / architecture owner
2MagistrTheOne author, HF namespace, runtime repo
3MINAKANUSHI / MINA architecture family / short name
I WILL SURVIVE. NULLXES.
