Generation architectures
These families are selected from the model’s model_type and routed to family-specific Java implementations.
| Family | model_type | Examples / variants | Status |
|---|---|---|---|
| Llama | llama | Llama, TinyLlama | First-class |
| Qwen2 | qwen2 | Qwen2 checkpoints | First-class |
| Qwen3 | qwen3 | 0.6B, 4B | First-class |
| Qwen3 MoE | qwen3_moe | Qwen3 MoE | First-class |
| Gemma 2 | gemma2 | Gemma 2 text | First-class |
| Gemma 3 Text | gemma3_text | Gemma 3 text | Help wanted |
| Gemma 4 | gemma4 | Gemma 4 text | Help wanted |
| Mistral | mistral | Mistral checkpoints | First-class |
| Mixtral | mixtral | Mixtral MoE | Supported |
| GPT-2 | gpt2 | GPT-2 | Supported |
| Granite Dense | granite | Granite 4.1 3B LLMON | Active support |
| Granite Hybrid | granitemoehybrid | Granite 4.0 Mamba/MoE | First-class |
| Diffusion Gemma | diffusion_gemma | Diffusion Gemma | Specialized |
| Nemotron Labs Diffusion | nemotron_labs_diffusion | Autoregressive and diffusion | Specialized |
Granite distinction: dense granite and granitemoehybrid are different architectures. They share some Granite concepts, but their layer implementations and tensor layouts are not interchangeable.
Embedding models
| Family | Usage | Status |
|---|---|---|
| BERT | Text embeddings | Supported |
| XLM-RoBERTa | Text embeddings | Supported |
| EmbeddingGemma 2 | Text embeddings, pooling, normalization, MRL truncation | Implemented; registration pending |
| Sentence-transformer / LEAF | Existing embedding workflow | Supported with parity limitations |
EmbeddingGemma 2 is currently a text-only implementation. The broader checkpoint may contain multimodal metadata, but Reckoning’s implemented path performs text encoding and embedding.
Classifiers
The documented classification path is intended primarily for encoder families:
- BERT: sequence classification such as jailbreak detection.
- XLM-RoBERTa: multilingual language classification.
The shared builder can technically attempt classification for other dispatched families, but those combinations are not generally validated as classifier checkpoints.
Tokenizers
The Grace tokenizer layer supports model-aware loading of:
- BERT and XLM-RoBERTa tokenizers
- Llama and Qwen-style tokenizers
- Gemma tokenizers
- Byte-level BPE and fast tokenizer JSON files
- Added tokens, special-token metadata, and chat templates
Tokenizer behavior is part of model support. BOS/EOS handling, reserved tokens, and chat_template.jinja can materially change model behavior.
Capabilities
| Capability | Coverage |
|---|---|
| JQ4 / Q4 quantization | Strongest support on causal generation models |
| Tool calls | Llama, Qwen, Gemma 4, Mixtral, and default parsers |
| Thinking output | Explicit Qwen3 reasoning parser |
| LLMON | Granite LLMON prompt DSL and Granite 4.1 support |
| Diffusion generation | Diffusion Gemma and Nemotron Labs Diffusion |
| Tensor providers | Native SIMD, Panama, and optional GPU paths |
Tool support is not purely an architecture property. It depends on the checkpoint’s chat template, tokenizer, output format, and tool-call parser. LLMON support is specific to LLMON-trained Granite checkpoints, not every Granite model.
How to read status
- First-class: normal dispatch, tests, and documented usage.
- Supported: working implementation with narrower validation.
- Specialized: separate runtime behavior such as diffusion.
- Active support: currently being ported or validated.
- Help wanted: code exists, but output correctness or parity is not reliable enough to call supported.
- Implemented; registration pending: model code exists but automatic dispatch is incomplete.