MedusaGemma-E4B

MedusaGemma-E4B

Tagline: One spine. Many heads.

Started as a MacBook Pro experiment and grew into a full composite release.

Model Summary

MedusaGemma-E4B is the fused raw release behind Pocket Polymath: a compact local model for reasoning, coding, and stack-assisted reliability.

Parameter note: This follows Google's E4B convention: approximately 4.5B effective parameters / ~8B including embeddings.

Built from the ground up on a single M5 Pro 48GB MacBook Pro, it was spliced, merged, dojo-trained, and reinforced into a sharper composite system.

What Went Into It

This is a single fused artifact, not a sidecar-heavy bundle.

It starts with google/gemma-4-E4B-it, then layers in donor influence, adapter passes, and reliability work so the final release is the creature plus the armor-ready base.

Primary lineage:

  • google/gemma-4-E4B-it
  • DavidAU/gemma-4-E4B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking
  • DavidAU/gemma-4-E4B-it-The-DECKARD-V2-Strong-HERETIC-UNCENSORED-Thinking
  • Jiunsong/supergemma4-e4b-abliterated

Stamsam custom dojo / training

The adaptation stack was built on:

  • pocket-polymath-seam-adapter
  • pocket-polymath-legacy-adapter
  • pocket-polymath-legacy-4096-adapter
  • pocket-polymath-broad-growth-adapter
  • pocket-polymath-ultimate-dojo-adapter

Synthetic dojo data came from glm-5.1:cloud via Ollama, plus Gemini-generated material from the same workflow.

Best For

  • local/offline assistant work
  • structured business reasoning
  • general writing and summarization
  • practical code generation
  • arithmetic-assisted answers with the arithmetic router
  • missing-data refusal with the refusal router
  • optional RC2.1 code repair in sandboxed workflows

Keep In Mind

  • The raw model can still make arithmetic mistakes without the stack.
  • The raw model can still miss exact refusal behavior without the stack.
  • The raw model can still miss strict formatting constraints without validators.
  • Code repair mode needs a sandbox and test harness.

Evaluation Snapshot

  • Fresh full holdout: 40/40
  • External-lite cleaned: 256/265, 96.6%
  • MBPP repair mode: 20/20 final on the cleaned slice and fresh holdout
  • Reasoning leaks: 0
  • Bad claims: 0

Safety and Reliability

The full Pocket Polymath stack adds:

  • reasoning leak guard
  • missing-data refusal router
  • arithmetic router
  • structured-output validators
  • completion checks
  • optional sandboxed code verifier/repair mode

Release Names

  • Raw model: MedusaGemma-E4B
  • MLX 4-bit: MedusaGemma-E4B-MLX-Q4
  • Default stack: Pocket Polymath: MedusaGemma-E4B RC2
  • Optional code mode: Pocket Polymath: MedusaGemma-E4B RC2.1 Code Repair
Downloads last month
26
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for stamsam/MedusaGemma-E4B

Finetuned
(316)
this model
Quantizations
1 model