hvm-gemma4/gemma4-26b-hvm4

Gemma 4 26B via Bend2/HVM4

Control-plane adapter for running the Gemma 4 26B A4B model through Bend 0.2.38 and HVM4 4.0.x before local Ollama inference.

prompt → Bend control generation → HVM4 control execution → Ollama/Gemma 4

HVM4 performs bounded pure control computation; model tensor inference remains in Ollama. This repository does not claim that Gemma tensor operations execute inside HVM4.

Model

The included Ollama Modelfile derives from gemma4:26b and creates the local gemma4:26b-hvm4 lane.

ollama create gemma4:26b-hvm4 -f hvm4/Modelfile.26b-hvm4
HVM_GEMMA_MODEL=gemma4:26b-hvm4 ./run-hvm4.sh "Reply exactly: HVM4_OK"

Compatibility

• Bend: exactly 0.2.38
• HVM4: 4.0.x
• Runtime invocation: generated HVM4 program with -C1
• Default local model: gemma4:26b-hvm4

Included files

• run-hvm4.sh: validated HVM4/Ollama runner
• hvm4/Main.hs: Haskell adapter
• hvm4/mailbox.bend: Bend2 mailbox classifier source
• hvm4/mailbox-bend2.sh: mailbox substrate entrypoint
• hvm4/Modelfile.26b-hvm4: Ollama model definition

License

Model usage remains subject to the Gemma Terms of Use. Adapter code follows the companion project license.