How harqer/iq works
IQ is an experimental hybrid reasoning language model, with a separate agent harness and a canonical weight-transfer package. Its main runtime path combines token-time Mamba state mixing, long-context attention, routed experts, and depth residuals, then supports multimodal input and optional iterative reasoning before language-model output and MTP. Transfer remaps donor capabilities into an IQ checkpoint; training and the harness provide adjacent workflows. Edges are limited to relationships stated in the README or visible in the supplied source index and excerpts; unsampled internals are not inferred.
Main entry points: hybrid.py Hybrid Backbone · runtime.py Agent Runtime · complete_transplant.py Weight Transfer · model.py IQ Language Model
Components 28
Hybrid Backbone
- IQ Language Model
iq_model/model.py - Hybrid Backbone
iq_model/hybrid.py - Mamba-3 MIMO
iq_model/state/mamba3.py - CSA/HCA Context
iq_model/attention/compressed.py - Dense Attention Control
iq_model/attention/context_dense.py - Stable LatentMoE
iq_model/mlp/latent_moe.py - SiTU Activation
iq_model/mlp/situ.py - Block AttnRes
iq_model/residual/attnres.py - mHC Ablation
iq_model/residual/mhc.py - Multimodal Pathway
iq_model/multimodal.py
Reasoning and Output
- Reasoning Recurrence
iq_model/reasoning/recurrence.py - Reasoning Verifier
iq_model/reasoning/verifier.py - Energy Critic
iq_model/energy/critic.py - MTP Objective
iq_model/objectives/mtp.py
Weight Transfer
- Mamba Checkpoint
- GLM-5.3 Weights
- Weight Transfer
iq_transfer/complete_transplant.py - Mamba Widening
iq_transfer/mamba3_direct.py - GLM Remapping
iq_transfer/glm53.py - IQ Checkpoint
iq_transfer/complete_transplant.py
Training and Harness
- Pretraining
iq_training/pretraining.py - Agent Runtime
iq_harness/runtime.py - Conversation Session
iq_harness/session.py - Agent Skills
iq_harness/skills.py - Agent Tools
iq_harness/tools.py - Model Adapter
iq_harness/model.py
Other components
- Model User
- Agent User
Connections 26
- Model User → IQ Language Model: requests generation
- IQ Language Model → Hybrid Backbone: runs backbone
- Hybrid Backbone → Mamba-3 MIMO: mixes token states
- Hybrid Backbone → CSA/HCA Context: addresses long context
- Hybrid Backbone → Stable LatentMoE: routes experts
- Stable LatentMoE → SiTU Activation: uses activation
iq_model/mlp/latent_moe.py - Hybrid Backbone → Block AttnRes: mixes depth residuals
- Hybrid Backbone → mHC Ablation: retains ablation
- Hybrid Backbone → Dense Attention Control: offers control path
- Hybrid Backbone → Multimodal Pathway: accepts visual input
iq_model/hybrid.py - Multimodal Pathway → Mamba-3 MIMO: processes visual stream
- Hybrid Backbone → Reasoning Recurrence: refines reasoning state
- Reasoning Recurrence → Reasoning Verifier: may request evidence
iq_model/reasoning/recurrence.py - Reasoning Verifier → Energy Critic: supports EBM backend
iq_model/energy/critic.py - Hybrid Backbone → MTP Objective: predicts future tokens
- Mamba Checkpoint → Mamba Widening: supplies foundation
- GLM-5.3 Weights → GLM Remapping: supplies capabilities
- Mamba Widening → Weight Transfer: widens weights
iq_transfer/complete_transplant.py - GLM Remapping → Weight Transfer: remaps capabilities
- Weight Transfer → IQ Checkpoint: assembles checkpoint
iq_transfer/complete_transplant.py - IQ Checkpoint → Pretraining: provides model weights
- Agent User → Agent Runtime: starts agent run
iq_harness/runtime.py - Agent Runtime → Conversation Session: loads and appends
iq_harness/runtime.py - Agent Runtime → Agent Skills: uses skills
iq_harness/runtime.py - Agent Runtime → Agent Tools: dispatches tools
iq_harness/runtime.py - Agent Runtime → Model Adapter: requests model turn
iq_harness/runtime.py