Substrate-Lens

Why does Qwen2.5 act so different under the hood than Llama-3?

A geometric profiler for any HuggingFace causal LM. Pick a model, see its architectural signature, per-layer anisotropy heatmap, recommended LoRA placement, and speculative-decoding suitability — all computed from the model's actual hidden-state geometry, not from architecture metadata.

Buy a license Try the free Space 
================================================================================
SUBSTRATE-LENS DIAGNOSTIC REPORT: Qwen/Qwen2.5-1.5B
================================================================================
>>> ARCHITECTURAL SIGNATURE: Type A: Extreme Highway (Cliff Exploder)
>>> SPECULATIVE DECODING SUITABILITY: HIGH (Score: 0.89)
    Recommended Anchor: Layer 25 (Last expressway junction)

LAYER MANIFOLD VISUALIZATION (Anisotropy Heatmap)
--------------------------------------------------------------------------------
L000 [0.10] ██░░░░░░░░░░░░░░░░░░ [DIVERSE]   Rank Target: 2/8 (Half)
L001 [1.00] ████████████████████ [HIGHWAY]   Rank Target: 3/8 (Half)
L002 [1.00] ████████████████████ [HIGHWAY]   Rank Target: 3/8 (Half)
...
L024 [0.99] ████████████████████ [HIGHWAY]   Rank Target: 3/8 (Half)
L025 [0.99] ████████████████████ [ANCHOR ]   Rank Target: 3/8 (Half)
L026 [0.13] ███░░░░░░░░░░░░░░░░░ [CLIFF  ]   Rank Target: 2/8 (Half)
L027 [0.37] ███████░░░░░░░░░░░░░ [FINAL  ]   Rank Target: 3/8 (Half)
--------------------------------------------------------------------------------

What you get

Architectural Signature

Classifies the model into one of three macro-shapes: Type A (extreme rank-1 highway with cliff fan-out, like Qwen2.5), Type B (gradual unpacker, like Llama / Pythia), or Type C (relaxed isotropic, common in small models).

LoRA Placement Score

Continuous per-layer S_LoRA score telling you which layers will actually accept fine-tuning weight. Comes with a copy-pasteable peft.LoraConfig(rank_pattern=...) dict.

Speculative Decoding Anchor

Identifies the optimal layer for mounting feature-extraction speculation heads (EAGLE-3, Medusa, Lookahead Decoding) based on the model's rank-1 expressway length.

Single binary, no Python required

Ship a 973 MB Linux ELF binary or a 2.7 GB Docker image. The algorithmic IP is C++ machine code — no source visible at runtime.

Works on any HF causal LM

12+ model families supported out of the box: Qwen2/2.5, Llama, Mistral, Phi-3, Gemma/Gemma2, Qwen2-MoE, Mixtral, GPT-NeoX, GPT-2, GPT-J, GPT-Neo. Pythonic duck-typed fallback covers most unrecognized architectures automatically.

JSON for CI/CD, Markdown for humans

Every invocation writes both a terminal-friendly ASCII dashboard and machine-readable JSON for downstream pipelines.

Pricing

One-time license, 3-machine activation, validates against Polar (built-in international VAT/tax handling).

Hobbyist

$19/month

or $190/year (save ~17%)

  • All features
  • 3 machines per key
  • Both CPU and CUDA builds
  • Community support
Buy Monthly — $19/mo Buy Yearly — $190/yr

Professional

$49/month

or $490/year (save ~17%)

  • Everything in Hobbyist
  • Priority email support
  • Early access to new architectures
  • Custom integration help
Buy Monthly — $49/mo Buy Yearly — $490/yr

Need a team/enterprise license, custom architecture support, or invoicing? Contact sales.

How it works

  1. Buy a license via Polar Checkout — payment + VAT/tax handled automatically.
  2. Receive your license key by email within minutes.
  3. Set SUBSTRATE_LENS_KEY as an environment variable, or pass --key on the CLI.
  4. Run substrate-lens <model_id> --output_dir ./reports.
  5. The binary validates your key against Polar on first run, then caches the validation locally (signed, tamper-resistant). Re-validates every 7 days.

The cached validation tolerates up to 30 days offline before requiring fresh network access. The binary never crashes on a bad license — it prints a clean error and exits with a code that's easy to handle in CI/CD pipelines.

Built for engineers who actually read the docs