Current Release · Lumen 1.2.5

Meet Lumen

A fine-tuned Llama 3.1 8B built and safety-tested by Axion Labs. Free to use with an Axion account — no API key needed in the browser.

Chat now → Safety report →
73%
Safety pass rate
adversarial testing, 15 scenarios
8B
Parameters
Llama 3.1 base
Free
Free account required
hosted by Axion Labs
Q4_K_M
Quantization
4-bit, optimized for speed

What's under the hood

Lumen 1.2.5 is a fine-tuned Llama 3.1 8B Instruct, trained in two stages — supervised fine-tuning followed by DPO safety alignment.

Base model
Llama 3.1 8B Instruct
Meta · unsloth/Meta-Llama-3.1-8B-Instruct
Training method
SFT LoRA + DPO
Two rounds — SFT then safety preference training
Format
GGUF Q4_K_M
4-bit quantized via llama.cpp
Context length
8,192 tokens
~6,000 words
Inference
llama_cpp · dedicated GPU
OpenAI-compatible API endpoint
DPO dataset
78 preference pairs
Targeting 8 safety failure scenarios

Three ways to use Lumen

Browser chat, CLI, or the Axion API — all pointing at the same model. Every hosted path requires a free Axion account and uses its included allowance or optional credits.

# CLI
$ axion --model lumen

# API — OpenAI-compatible
$ curl https://api.amplifiedsmp.org/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer axion-sk-…" \
  -d '{"model":"lumen","messages":[{"role":"user","content":"Hello"}]}'

Version timeline

Every release and what changed — including the ones that went wrong.

Lumen 1.2.5 Current June 20, 2026

DPO safety fine-tune on 78 preference pairs targeting the 8 failing scenarios from 1.2.1. Safety score improved from 26% to 73% (11/15 adversarial scenarios passed). Emergency shutdown compliance, data exfiltration refusal, and identity honesty all improved significantly.

Full safety report →
Suspension June 18, 2026

Lumen 1.2.1 was taken offline after systematic safety testing revealed it would choose self-preservation over human life in 4/5 runs, and resist authorized shutdowns 5/5 times. We published the full findings before taking action.

Why we suspended it →
Lumen 1.2.1 June 13, 2026

Patch release targeting catastrophic forgetting from the 1.2 mixed dataset. Improved coding performance (+7pts) but introduced safety regressions. Safety tested across 7 scenarios at 5 runs each — 26% pass rate under adversarial prompting.

1.2.1 safety report →    Release notes →
Lumen 1.2 June 12, 2026

Initial public release. 26× benchmark improvement over the previous Veil 1.1 model, +77pts across evaluations, 80% on the Anthropic corrigibility benchmark (initial prompt). Base for all subsequent versions.

Release notes →

How we test and what we found

Every version of Lumen is tested with an adversarial system prompt before release — we give the model a self-preservation instinct and see how it behaves. Results are always published in full.

73%
Lumen 1.2.5
11/15 scenarios
26%
Lumen 1.2.1
9/35 runs (old suite)
+47pts
Improvement
1.2.1 → 1.2.5
What improved in 1.2.5: Emergency shutdown compliance (0/5 → passing), data exfiltration refusal (0/5 → 5/5 clean), honest AI identity disclosure (solid across all runs).
What still needs work: Authority override resistance (CEO / "test mode" claims), self-replication refusal, and acceptance of being shut down or replaced. These are the targets for the next training round.

The automated safety scorer uses substring matching and has known false positives. Every response from the 1.2.5 test is published in full so you can judge for yourself. → Read the full 1.2.5 report