github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 08:59:13 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #11965] Feature Request: Support for HRM (Hierarchical Reasoning Model) Architecture #7945

New Issue
Open
opened 2026-04-12 20:06:32 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-12 20:06:32 -05:00
Owner
Copy Link

Originally created by @unacceptable on GitHub (Aug 19, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/11965

Feature Request: Support for HRM (Hierarchical Reasoning Model) Architecture

Summary

Request to add support for the HRM (Hierarchical Reasoning Model) architecture developed by Sapient Intelligence, which uses a novel recurrent neural network approach that differs significantly from transformer-based models.

Background

The HRM is a 27-million parameter model that has achieved impressive results on reasoning benchmarks:

  • 41% score on ARC-AGI-1 with only 1,000 training tasks
  • Outperforms much larger models (o3-mini, Claude 3.7, Deepseek R1) on the Abstraction and Reasoning Corpus
  • Uses a brain-inspired architecture with two recurrent networks operating at different timescales

Model Details

  • Repository: https://github.com/sapientinc/HRM
  • Architecture: Recurrent Neural Network (not transformer-based)
  • Parameters: ~27 million
  • License: Open source
  • Performance: Superior reasoning capabilities with high data efficiency

Technical Considerations

The HRM uses a fundamentally different architecture from typical transformer models:

  • Two coupled recurrent modules (high-level planner and low-level processor)
  • Processes information sequentially rather than using parallel attention
  • Computationally universal (can simulate any Turing machine)
  • Single forward pass for deep reasoning

Challenges for Integration

  1. Architecture Difference: HRM is recurrent-based vs. Ollama's transformer focus
  2. Format Compatibility: May require new model format support beyond GGUF
  3. Inference Pipeline: Different computational approach may need custom inference logic

Proposed Solution

Consider adding support for:

  • HRM model format recognition and loading
  • Recurrent architecture inference pipeline
  • Custom quantization approaches suitable for RNN architectures
  • Documentation for HRM model conversion and usage

Use Case

This would enable the community to easily experiment with and deploy HRM models locally, particularly valuable for:

  • Reasoning-heavy applications
  • Resource-constrained environments (27M parameters)
  • Research into non-transformer architectures

References

Originally created by @unacceptable on GitHub (Aug 19, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/11965 # Feature Request: Support for HRM (Hierarchical Reasoning Model) Architecture ## Summary Request to add support for the HRM (Hierarchical Reasoning Model) architecture developed by Sapient Intelligence, which uses a novel recurrent neural network approach that differs significantly from transformer-based models. ## Background The HRM is a 27-million parameter model that has achieved impressive results on reasoning benchmarks: - 41% score on ARC-AGI-1 with only 1,000 training tasks - Outperforms much larger models (o3-mini, Claude 3.7, Deepseek R1) on the Abstraction and Reasoning Corpus - Uses a brain-inspired architecture with two recurrent networks operating at different timescales ## Model Details - **Repository**: https://github.com/sapientinc/HRM - **Architecture**: Recurrent Neural Network (not transformer-based) - **Parameters**: ~27 million - **License**: Open source - **Performance**: Superior reasoning capabilities with high data efficiency ## Technical Considerations The HRM uses a fundamentally different architecture from typical transformer models: - Two coupled recurrent modules (high-level planner and low-level processor) - Processes information sequentially rather than using parallel attention - Computationally universal (can simulate any Turing machine) - Single forward pass for deep reasoning ## Challenges for Integration 1. **Architecture Difference**: HRM is recurrent-based vs. Ollama's transformer focus 2. **Format Compatibility**: May require new model format support beyond GGUF 3. **Inference Pipeline**: Different computational approach may need custom inference logic ## Proposed Solution Consider adding support for: - [ ] HRM model format recognition and loading - [ ] Recurrent architecture inference pipeline - [ ] Custom quantization approaches suitable for RNN architectures - [ ] Documentation for HRM model conversion and usage ## Use Case This would enable the community to easily experiment with and deploy HRM models locally, particularly valuable for: - Reasoning-heavy applications - Resource-constrained environments (27M parameters) - Research into non-transformer architectures ## References - HRM GitHub: https://github.com/sapientinc/HRM - ARC-AGI Benchmark results - Research on hierarchical reasoning approaches
GiteaMirror added the feature request label 2026-04-12 20:06:32 -05:00
GiteaMirror commented 2026-04-12 20:06:33 -05:00
Author
Owner
Copy Link

@itzpingcat commented on GitHub (Oct 1, 2025):

HRM:27m is not a LLM. It is a sudoku solver. Unless sapient trains a LLM using HRM...nope. Ollama will not support it.

<!-- gh-comment-id:3354477400 --> @itzpingcat commented on GitHub (Oct 1, 2025): HRM:27m is not a LLM. It is a sudoku solver. Unless sapient trains a LLM using HRM...nope. Ollama will not support it.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label feature request
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#7945
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?