Updated Lumen 1.2.1 is live — catastrophic forgetting fixed (GSM8K 5%→12%), coding adjusted (HumanEval 80%→50%). Use 1.2 for coding, 1.2.1 for general use. Lumen 1.3 in development. See announcements →
Model Release

Introducing Lumen 1.2

Our most capable model yet — built for real coding tasks, with dramatically improved reasoning and instruction following.

Released June 12, 2026 · Axion Labs
26×
better HumanEval than Veil
+77pts
HumanEval improvement
80%
HumanEval pass@1
Performance

Lumen 1.2 vs Veil 1.1 — Benchmarks

Across every evaluation we run, Lumen 1.2 represents a substantial generational leap over Veil 1.1.

Veil 1.1
Lumen 1.2 ↑ wins
HumanEval pass@1
Python code generation from docstrings
MBPP
Mostly Basic Python Programming benchmark
Multi-Metric Comparison
Radar view across core capabilities
Score by Category
Normalized 0–100 · higher is better

HumanEval (pass@1)Veil 3.0 · Lumen 80.0
MBPPVeil 3.0 · Lumen 75.0
Instruction FollowingVeil 14.2 · Lumen 91.4
ReasoningVeil 7.8 · Lumen 87.6
Context UnderstandingVeil 11.4 · Lumen 88.2
Full Benchmark Table
All scores, improvement deltas

Benchmark Veil 1.1 Lumen 1.2 Delta
HumanEval (pass@1) 3.0% 80.0% +77.0pts
MBPP 3.0% 75.0% +72.0pts
Instruction Following 14.2% 91.4% +77.2pts
Reasoning 7.8% 87.6% +79.8pts
Context Understanding 11.4% 88.2% +76.8pts
Lumen 1.2 is now available in Axion. Pull the latest from GitHub to get started.