The lab · live

Canada's foundation model.

1.34 billion parameters. Trained from scratch on Apple Silicon. release-v1 locked.

OPS_CONSOLE·edge-2-large · release-v1·build: 8a5b004 → d242103·node: jl1.local · 64 GB · MLXno-data
MODEL
nameEdge-2 Large
params1,191,282,688
archRoPE · SwiGLU · GQA · BPE-32K
statusconverged
releaserelease-v1
TRAINING
best_val3.3518
best_step56,000
last_val3.6770
loss2.9900
throughput511 tok/s
ATTESTATION
sigML-DSA-87 · FIPS 204 L5
chain~/.axe/knox/audit/chain.jsonl
watchdog5min · local + external
cascadefail-closed · quarantine
hashsigned · auditor-replayable
NEXT_PHASE
nameSFT v3 — Canadian domain
pairs16,900
tokens_est8.6M
base_ckptrelease-v1
statusqueued
LIVE_COMMAND
$ python3 train_mlx.py \
  --config large-v2 \
  --train-tokens master_blend_stage2_32k.npy \
  --val-tokens axe_mixed_val_32k.npy \
  --tokenizer tokenizer_bpe_32k.json \
  --output checkpoints/edge2-large-v2 \
  --steps 100000 \
  --resume checkpoints/edge2-large-v2/release-v1
FLEET
NOVAPrimary Training Cluster
FORGECommand & Ops
VIGILSecurity & Evaluation
GHOSTStealth Operations
REAPEREdge Processing

Edge-2 Large · the numbers.

1.34B
Parameters
exact 1191.3M
3.3518
Best val_loss
at step 56,000
80,010
Final pretraining step
stopped 2026-05-12 (converged · LR-decayed plateau)
0
Foreign weights
zero fine-tune lineage from foreign models
BPE-32K
Tokenizer
Canadian-tuned vocabulary
Commercial
License
weights AXE-owned · deployment customer-portable
Current release

release-v1 — step 56,000 — locked.

Model
Edge-2 Large
Parameters
1,191,282,688
Architecture
RoPE · SwiGLU · GQA · BPE-32K
Training hardware
JL1 · Apple Silicon · MLX
Best val_loss
3.3518 @ step 56,000
Release symlink
release-v1 → step_0056000
Converged
Tue, 12 May 2026 17:35:00 GMT
Status
● converged
● VERIFIED
Knox EMA Attestation
algorithm: ML-DSA-87 (FIPS 204 L5)
chain_path: ~/.axe/knox/audit/chain.jsonl
bundle: knox_v3/ema/edge_attestation.py
watchdog: knox_watchdog_attestation_validity.py
cadence: 1h local · boot preemptive
fail_mode: CRITICAL = quarantine + W2/W3 cascade
fleet_signed_by: JL1 · JL2 · JL3
Every release-v1 inference response is attestable: model weights hashed, signed with the fleet KMS, and chain-linked to the audit-chain root. Auditors can replay the chain offline.
Next phase
SFT v3 — Canadian domain · base checkpoint release-v1 · 16,900 pairs · 8.6M tokens · status queued.
Plan: ~/Desktop/axiom/axe-dl/training/NEXT_BLEND_PLAN.md. SFT trigger gate (val_loss < 3.5) already met at step 56,000.
Edge roadmap

812K → 11B. Each step shipped.

Edge-1 (V1)
812K
done
Edge-2 Medium (V1)
85M
done
Edge-2 Small V2
15.7M
done
Edge-2 Large V2
1.34B
→ release-v1
converged · release-v1
Edge-2 Large SFT v3
1.34B
queued
Edge-2 Large Stage 3
1.34B
planned
Edge-3 7B
7B
planned
Edge-3 11B
11B
planned
The fleet

Five nodes. Five callsigns. One Canadian mesh.

NOVAstandby
Primary Training Cluster
TOR-1
FORGEonline
Command & Ops
TOR-2
VIGILonline
Security & Evaluation
TOR-3
GHOSTstandby
Stealth Operations
TOR-4
REAPERstandby
Edge Processing
TOR-5

We don't ship a loss curve. We ship the model.

Lab beta access · Knox-signed

Apply for early Edge-2 Large access.

Tell us about your deployment context. Every submission is signed with HMAC-SHA256 and appended to the Knox audit chain — non- repudiable, auditable, retainable. We respond within 5 business days.

POST → /api/lab-apply · HMAC-SHA256 · Knox chain.jsonl