Supported model families

llamadart has no model allowlist. Any GGUF that the pinned llama.cpp runtime can load will generate text. What varies by family is how chat messages, tool calls and reasoning are formatted and parsed, and that follows from the model's chat template.

How a family is recognized#

• GGUF: llamadart reads the chat template embedded in the file (tokenizer.chat_template, or tokenizer.chat_template.tool_use when tools are passed) and detects its format from marker strings, in llama.cpp's priority order. It does not look at the model name or architecture.
• .litertlm: llamadart picks a template from the file name (gemma-4, gemma-3n, gemma-3/gemma-2, qwen3 or qwen2.5) for format detection and output parsing. Set ModelParams.chatTemplate to override it. On native LiteRT-LM, getMetadata() reports the template the runtime renders as tokenizer.chat_template, and the parsing template, when it differs, as llamadart.chat_template.parsing. A bundle with no matching family and no override reports no template.
• No match: the template renders as-is, with no tool-call or reasoning parsing. If you pass tools, llamadart adds a generic JSON tool-call instruction instead.

To see what a model will receive, render the prompt with engine.chatTemplate(...); see Chat templates and output parsing. For native .litertlm the prompt comes from the template the runtime renders for text-only requests: the file-name template for Qwen3 bundles, and the bundle's embedded template for the others, such as Qwen3.5 and Gemma. Dart renders it, so it shows what that template produces, not the runtime's exact bytes; requests with media always use the embedded template. The other result fields, such as format, still come from the parsing template.

Evidence levels#

• Validated: the llamadart validation pack (packages/llamadart_validation) has a profile that runs the real model.
• Template-tested: CI detects and renders the family's published chat template, and parses sample output where one is recorded. Templates come from llama.cpp's template set (release b10549) and from real GGUF files.

Chat families#

Other families whose published templates are template-tested with their own format: GPT-OSS, Seed-OSS, Nemotron Nano v2, Apertus, Kimi K2 and K3, MiniMax M1, M2 and M3, Command R7B, Cohere2 MoE, MiniCPM5, Hunyuan Hy3, Solar Open, Poolside Laguna, Muse Glimmer, Functionary v3.1 and v3.2, and FireFunction v2. Nemotron 3 Nano and StepFun 3.5 Flash use the Qwen3-Coder XML format; Hermes 2 Pro, Hermes 3, Bielik, Reka Edge, MiMo-VL and Apriel 1.5 use the Hermes format.

"Generic JSON tool calls" means llamadart supplies a fallback protocol when the selected template has no recognized tool-call format: it asks the model to answer with a tool_call or response JSON object and parses that. This is an intentional extension to the tested llama.cpp behavior, not evidence of native tool support or reliable tool selection. Templates can omit tool definitions even when a grammar constrains the output. See Generic JSON fallback for the protocol, upstream comparison and prompt limitations.

Where a format defines reasoning markers, such as <think>, Gemma 4's thought channel or Mistral's [THINK], the parser returns reasoning separately from the reply. See Text generation and streaming.

Qwen3.5 .litertlm bundles render their own embedded chat template (#1134), so ModelParams.chatTemplate is ignored for these files. The Qwen3.5-0.8B_int8, Qwen3.5-0.8B-VL_int8 and Qwen3.5-0.8B_int8_gpu_composite_v0190 bundles (LiteRT-LM 0.19 and earlier) close an empty <think> block in the generation prompt and never read enable_thinking, so enableThinking: true produces no reasoning there. The Qwen3.5-0.8B_dynamic_wi8_afp32 bundle reads enable_thinking but needs LiteRT-LM 0.20.0 or later, so it fails on the pinned 0.19.1 runtime. Use the GGUF for Qwen3.5 thinking; Qwen3 and Gemma 4 .litertlm bundles support it.

Vision and audio#

GGUF vision and audio need a matching projector (mmproj) file. Gemma 4 E2B (vision and audio), Qwen3.5 (vision) and LFM2-VL (vision) are the families the examples use. .litertlm bundles process media themselves. See Multimodal input.

Speech, embeddings and decision models#

See Speech to text, Text to speech, Embeddings and Decision models.

If a family behaves oddly#

1. Render the prompt with engine.chatTemplate(...) and compare it with the model card.
2. Check which format was detected: ChatTemplateEngine.detectFormat(template).
3. For tool calls, try a family listed above as validated before debugging your own prompts.
4. Report the model repo and file in a GitHub issue.