[0000] inference.route(model=edge-2-large-1.34B, stream=true)
[0000] castle.dispatch(node=jl1, task=sprint, status=ok)
[0001] inference.route(model=edge-2-large-1.34B, stream=true)
[0001] castle.dispatch(node=jl1, task=sprint, status=ok)
[0002] inference.route(model=edge-2-large-1.34B, stream=true)
[0002] castle.dispatch(node=jl1, task=sprint, status=ok)
[0003] inference.route(model=edge-2-large-1.34B, stream=true)
[0003] castle.dispatch(node=jl1, task=sprint, status=ok)
[0004] inference.route(model=edge-2-large-1.34B, stream=true)
[0004] castle.dispatch(node=jl1, task=sprint, status=ok)
[0005] inference.route(model=edge-2-large-1.34B, stream=true)
[0005] castle.dispatch(node=jl1, task=sprint, status=ok)
[0006] inference.route(model=edge-2-large-1.34B, stream=true)
[0006] castle.dispatch(node=jl1, task=sprint, status=ok)
[0007] inference.route(model=edge-2-large-1.34B, stream=true)
[0007] castle.dispatch(node=jl1, task=sprint, status=ok)
[0008] inference.route(model=edge-2-large-1.34B, stream=true)
[0008] castle.dispatch(node=jl1, task=sprint, status=ok)
[0009] inference.route(model=edge-2-large-1.34B, stream=true)
[0009] castle.dispatch(node=jl1, task=sprint, status=ok)
[0010] inference.route(model=edge-2-large-1.34B, stream=true)
[0010] castle.dispatch(node=jl1, task=sprint, status=ok)
[0011] inference.route(model=edge-2-large-1.34B, stream=true)
[0011] castle.dispatch(node=jl1, task=sprint, status=ok)
[0012] inference.route(model=edge-2-large-1.34B, stream=true)
[0012] castle.dispatch(node=jl1, task=sprint, status=ok)
[0013] inference.route(model=edge-2-large-1.34B, stream=true)
[0013] castle.dispatch(node=jl1, task=sprint, status=ok)
[0014] inference.route(model=edge-2-large-1.34B, stream=true)
[0014] castle.dispatch(node=jl1, task=sprint, status=ok)
[0015] inference.route(model=edge-2-large-1.34B, stream=true)
[0015] castle.dispatch(node=jl1, task=sprint, status=ok)
[0016] inference.route(model=edge-2-large-1.34B, stream=true)
[0016] castle.dispatch(node=jl1, task=sprint, status=ok)
[0017] inference.route(model=edge-2-large-1.34B, stream=true)
[0017] castle.dispatch(node=jl1, task=sprint, status=ok)
[0018] inference.route(model=edge-2-large-1.34B, stream=true)
[0018] castle.dispatch(node=jl1, task=sprint, status=ok)
[0019] inference.route(model=edge-2-large-1.34B, stream=true)
[0019] castle.dispatch(node=jl1, task=sprint, status=ok)
[0020] inference.route(model=edge-2-large-1.34B, stream=true)
[0020] castle.dispatch(node=jl1, task=sprint, status=ok)
[0021] inference.route(model=edge-2-large-1.34B, stream=true)
[0021] castle.dispatch(node=jl1, task=sprint, status=ok)
[0022] inference.route(model=edge-2-large-1.34B, stream=true)
[0022] castle.dispatch(node=jl1, task=sprint, status=ok)
[0023] inference.route(model=edge-2-large-1.34B, stream=true)
[0023] castle.dispatch(node=jl1, task=sprint, status=ok)
[0024] inference.route(model=edge-2-large-1.34B, stream=true)
[0024] castle.dispatch(node=jl1, task=sprint, status=ok)
[0025] inference.route(model=edge-2-large-1.34B, stream=true)
[0025] castle.dispatch(node=jl1, task=sprint, status=ok)
[0026] inference.route(model=edge-2-large-1.34B, stream=true)
[0026] castle.dispatch(node=jl1, task=sprint, status=ok)
[0027] inference.route(model=edge-2-large-1.34B, stream=true)
[0027] castle.dispatch(node=jl1, task=sprint, status=ok)
[0028] inference.route(model=edge-2-large-1.34B, stream=true)
[0028] castle.dispatch(node=jl1, task=sprint, status=ok)
[0029] inference.route(model=edge-2-large-1.34B, stream=true)
[0029] castle.dispatch(node=jl1, task=sprint, status=ok)
[0030] inference.route(model=edge-2-large-1.34B, stream=true)
[0030] castle.dispatch(node=jl1, task=sprint, status=ok)
[0031] inference.route(model=edge-2-large-1.34B, stream=true)
[0031] castle.dispatch(node=jl1, task=sprint, status=ok)
[0032] inference.route(model=edge-2-large-1.34B, stream=true)
[0032] castle.dispatch(node=jl1, task=sprint, status=ok)
[0033] inference.route(model=edge-2-large-1.34B, stream=true)
[0033] castle.dispatch(node=jl1, task=sprint, status=ok)
[0034] inference.route(model=edge-2-large-1.34B, stream=true)
[0034] castle.dispatch(node=jl1, task=sprint, status=ok)
[0035] inference.route(model=edge-2-large-1.34B, stream=true)
[0035] castle.dispatch(node=jl1, task=sprint, status=ok)
[0036] inference.route(model=edge-2-large-1.34B, stream=true)
[0036] castle.dispatch(node=jl1, task=sprint, status=ok)
[0037] inference.route(model=edge-2-large-1.34B, stream=true)
[0037] castle.dispatch(node=jl1, task=sprint, status=ok)
[0038] inference.route(model=edge-2-large-1.34B, stream=true)
[0038] castle.dispatch(node=jl1, task=sprint, status=ok)
[0039] inference.route(model=edge-2-large-1.34B, stream=true)
[0039] castle.dispatch(node=jl1, task=sprint, status=ok)
[0040] inference.route(model=edge-2-large-1.34B, stream=true)
[0040] castle.dispatch(node=jl1, task=sprint, status=ok)
[0041] inference.route(model=edge-2-large-1.34B, stream=true)
[0041] castle.dispatch(node=jl1, task=sprint, status=ok)
[0042] inference.route(model=edge-2-large-1.34B, stream=true)
[0042] castle.dispatch(node=jl1, task=sprint, status=ok)
[0043] inference.route(model=edge-2-large-1.34B, stream=true)
[0043] castle.dispatch(node=jl1, task=sprint, status=ok)
[0044] inference.route(model=edge-2-large-1.34B, stream=true)
[0044] castle.dispatch(node=jl1, task=sprint, status=ok)
[0045] inference.route(model=edge-2-large-1.34B, stream=true)
[0045] castle.dispatch(node=jl1, task=sprint, status=ok)
[0046] inference.route(model=edge-2-large-1.34B, stream=true)
[0046] castle.dispatch(node=jl1, task=sprint, status=ok)
[0047] inference.route(model=edge-2-large-1.34B, stream=true)
[0047] castle.dispatch(node=jl1, task=sprint, status=ok)
[0048] inference.route(model=edge-2-large-1.34B, stream=true)
[0048] castle.dispatch(node=jl1, task=sprint, status=ok)
[0049] inference.route(model=edge-2-large-1.34B, stream=true)
[0049] castle.dispatch(node=jl1, task=sprint, status=ok)
[0050] inference.route(model=edge-2-large-1.34B, stream=true)
[0050] castle.dispatch(node=jl1, task=sprint, status=ok)
[0051] inference.route(model=edge-2-large-1.34B, stream=true)
[0051] castle.dispatch(node=jl1, task=sprint, status=ok)
[0052] inference.route(model=edge-2-large-1.34B, stream=true)
[0052] castle.dispatch(node=jl1, task=sprint, status=ok)
[0053] inference.route(model=edge-2-large-1.34B, stream=true)
[0053] castle.dispatch(node=jl1, task=sprint, status=ok)
[0054] inference.route(model=edge-2-large-1.34B, stream=true)
[0054] castle.dispatch(node=jl1, task=sprint, status=ok)
[0055] inference.route(model=edge-2-large-1.34B, stream=true)
[0055] castle.dispatch(node=jl1, task=sprint, status=ok)
[0056] inference.route(model=edge-2-large-1.34B, stream=true)
[0056] castle.dispatch(node=jl1, task=sprint, status=ok)
[0057] inference.route(model=edge-2-large-1.34B, stream=true)
[0057] castle.dispatch(node=jl1, task=sprint, status=ok)
[0058] inference.route(model=edge-2-large-1.34B, stream=true)
[0058] castle.dispatch(node=jl1, task=sprint, status=ok)
[0059] inference.route(model=edge-2-large-1.34B, stream=true)
[0059] castle.dispatch(node=jl1, task=sprint, status=ok)
[0060] inference.route(model=edge-2-large-1.34B, stream=true)
[0060] castle.dispatch(node=jl1, task=sprint, status=ok)
[0061] inference.route(model=edge-2-large-1.34B, stream=true)
[0061] castle.dispatch(node=jl1, task=sprint, status=ok)
[0062] inference.route(model=edge-2-large-1.34B, stream=true)
[0062] castle.dispatch(node=jl1, task=sprint, status=ok)
[0063] inference.route(model=edge-2-large-1.34B, stream=true)
[0063] castle.dispatch(node=jl1, task=sprint, status=ok)
[0064] inference.route(model=edge-2-large-1.34B, stream=true)
[0064] castle.dispatch(node=jl1, task=sprint, status=ok)
[0065] inference.route(model=edge-2-large-1.34B, stream=true)
[0065] castle.dispatch(node=jl1, task=sprint, status=ok)
[0066] inference.route(model=edge-2-large-1.34B, stream=true)
[0066] castle.dispatch(node=jl1, task=sprint, status=ok)
[0067] inference.route(model=edge-2-large-1.34B, stream=true)
[0067] castle.dispatch(node=jl1, task=sprint, status=ok)
[0068] inference.route(model=edge-2-large-1.34B, stream=true)
[0068] castle.dispatch(node=jl1, task=sprint, status=ok)
[0069] inference.route(model=edge-2-large-1.34B, stream=true)
[0069] castle.dispatch(node=jl1, task=sprint, status=ok)
[0070] inference.route(model=edge-2-large-1.34B, stream=true)
[0070] castle.dispatch(node=jl1, task=sprint, status=ok)
[0071] inference.route(model=edge-2-large-1.34B, stream=true)
[0071] castle.dispatch(node=jl1, task=sprint, status=ok)
[0072] inference.route(model=edge-2-large-1.34B, stream=true)
[0072] castle.dispatch(node=jl1, task=sprint, status=ok)
[0073] inference.route(model=edge-2-large-1.34B, stream=true)
[0073] castle.dispatch(node=jl1, task=sprint, status=ok)
[0074] inference.route(model=edge-2-large-1.34B, stream=true)
[0074] castle.dispatch(node=jl1, task=sprint, status=ok)
[0075] inference.route(model=edge-2-large-1.34B, stream=true)
[0075] castle.dispatch(node=jl1, task=sprint, status=ok)
[0076] inference.route(model=edge-2-large-1.34B, stream=true)
[0076] castle.dispatch(node=jl1, task=sprint, status=ok)
[0077] inference.route(model=edge-2-large-1.34B, stream=true)
[0077] castle.dispatch(node=jl1, task=sprint, status=ok)
[0078] inference.route(model=edge-2-large-1.34B, stream=true)
[0078] castle.dispatch(node=jl1, task=sprint, status=ok)
[0079] inference.route(model=edge-2-large-1.34B, stream=true)
[0079] castle.dispatch(node=jl1, task=sprint, status=ok)

Edge-2 Large · 1.34B · live

AI that belongs to Canada.

Trained from scratch. Deployed on your iron. Audit-grade by default.

Request Access

EDGE-2 LARGE · STATUS
params .................. 1,191,282,688
tokenizer ............... BPE-32K
train_corpus ............ stage-2 mixed
best_val ................ 3.3518
release ................. release-v1 (step 56k)
foreign_weights ......... 0
hardware ................ JL1 · Apple Silicon
↗ converged · 2026-05-12

As deployedrelease|release-v1step|56,000val_loss|3.3518attest|ML-DSA-87fleet|5 nodescanadian|100%2026-05-12T17:35Z

Trained from scratch

Parameters

1.34 billion

On Apple Silicon. In Canada. On hardware you can touch. Not a fine-tune of a foreign base — the architecture, the tokenizer, and every weight are ours.

A Letter from Our Founders

James Lewis, CEO
Michael Mayo, COO

Our platform is deployed across government, defense, and enterprise organizations that require complete data ownership and AI capability without cloud dependencies. We build it because the alternative — sending Canadian queries to American servers for processing by American weights under American jurisdiction — isn't ownership. It's rent.

AXE is built in Canada, deploys on hardware you own, and licenses the model to you commercially with a Knox audit chain underneath every inference. The weights stay AXE-owned · the deployment is yours to keep · the data never leaves your perimeter.

What the numbers actually say.

0.00B

Edge parameters

trained on Apple Silicon · best_val 3.3518

Canadians served

via Mountie Service Canada integration

External API spend

local inference · llama-server + Ollama

Zero

Foreign weights

no fine-tunes of foreign base models

<0ms

P95 inference latency

JL1 Apple Silicon · BPE-32K tokenizer

Yours to deploy

Edge licensing model

commercial license · runs on your iron · weights stay AXE-owned

Platform · five pillars

One stack. Five platform layers. Every layer Canadian.

Pillar 01 · CASTLE

Autonomous orchestration.

Autonomous sprint automation across the AXE fleet. Routes tasks to the best Canadian machine + model. Zero paid API spend.

Pillar 02 · Edge

The foundation model.

Native language model trained from scratch on Apple Silicon via MLX. Not a fine-tune. 1.34B parameters today; 11B by Q3 2027.

Pillar 03 · Algorithm

Recall and ranking.

Cuckoo-Sharded Recall (CSR) — collision-free hashmap + HNSW serving recommendation, oracle retrieval, and per-customer memory.

Pillar 04 · Knox

Procurement-grade security.

Post-quantum signing (ML-DSA-87 FIPS 204 Level 5). Audit chain, revocation freshness, key rotation. Fail-closed.

Pillar 05 · Supercharge

The cognitive scaffold.

Six post-training components wrap Edge-11B for 70B-class performance on regulated narrow tasks. Direct-sales only.

Ownership isn't a feature you can toggle on.
It's the architecture.

Two ways to deploy AI in Canada.

Traditional AI

Your data on their servers.
Their weights. Their rules.

↘ Queries leave Canadian soil
↘ Foreign jurisdiction
↘ Metered API costs forever
↘ Black-box weights
↘ Vendor lock-in

AXE

Your hardware. Your models.
Your rules. Zero external dependencies.

↗ Canadian inference. Not Virginia. Toronto.
↗ PIPEDA-ready · data residency
↗ $0 external API spend
↗ Commercial license · weights stay AXE-owned · deployment yours
↗ Protected B ready · FIPS 140-2 path

Pillar 02 · live · 1.34B parameters

Canada's foundation model. Trained now.

Edge-2 Large 1.34B converged on 2026-05-12 with best_val 3.3518. Trained from scratch on Apple Silicon — no foreign base model, no fine-tune lineage. The next step is supervised fine-tuning on 16,900 Canadian-domain pairs, then stage-3 pretrain at 3B tokens. By Q3 2027, Edge reaches 11B parameters and ships as the first household-name Canadian foundation model.

See the Lab ↗

Parameters

1,191,282,688

Architecture

RoPE · SwiGLU · GQA

Tokenizer

BPE-32K

Training hardware

JL1 · Apple Silicon · MLX

Best val_loss

3.3518 at step 56,000

Roadmap

Edge-3 11B by Q3 2027

Founder note

“If you're not invited to the table, you're probably on the menu. We built AXE because every other Canadian organization with serious AI ambitions was being told what to want by a US salesperson.”

— James Lewis, CEO

⨉

Bring your data home.

We deploy on your iron, integrate with your perimeter, and run audit-grade inference behind your firewall. Government, defense, and enterprise organizations running classified workloads ship with AXE.

Request Access or see the Lab ↗