Model type
Language model
20 Downloads
Capabilities
Minimum system memory
Tags
Model type
Language model
20 Downloads
Capabilities
Model type
Language model
20 Downloads
Capabilities
Minimum system memory
Tags
Model type
Language model
20 Downloads
Capabilities
Minimum system memory
Tags
Last updated
Updated 16 days agobyREADME
This catalog groups two native MLX downloads of the Heretic v2 language checkpoint. It declares Thinking, Vision and Tool Use and provides an Enable Thinking on/off control, enabled by default.
| Download | Weight file | Vision storage |
|---|---|---|
| Original | 8.60 GB | Original FP16 vision tower |
| Compact | 8.41 GB | 22 of 27 vision MLP expansion layers quantized to native 2-bit; the five most sensitive blocks kept FP16, remaining vision tensors preserved |
All 2,057 language tensors are identical between variants, including the original Hadamard representation. Compact saves about 187 MB of disk and loaded parameter allocation. Its vision quantization is lossy. The original is the first catalog option and preserves the original vision weights.
The Original and Compact variants have the same suggested unified-memory tiers.
| Mac RAM | Guidance |
|---|---|
| 8 GB | Not practical for either variant |
Custom Fields
Special features defined by the model author
Enable Thinking
: boolean
(default=true)
Enables reasoning before the answer. Turn off for a direct answer.
Parameters
Custom configuration options included with this model
Sources
The underlying model files this model uses
Minimum system memory
Tags
Last updated
Updated 16 days agobyREADME
This catalog groups two native MLX downloads of the Heretic v2 language checkpoint. It declares Thinking, Vision and Tool Use and provides an Enable Thinking on/off control, enabled by default.
| Download | Weight file | Vision storage |
|---|---|---|
| Original | 8.60 GB | Original FP16 vision tower |
| Compact | 8.41 GB | 22 of 27 vision MLP expansion layers quantized to native 2-bit; the five most sensitive blocks kept FP16, remaining vision tensors preserved |
All 2,057 language tensors are identical between variants, including the original Hadamard representation. Compact saves about 187 MB of disk and loaded parameter allocation. Its vision quantization is lossy. The original is the first catalog option and preserves the original vision weights.
The Original and Compact variants have the same suggested unified-memory tiers.
| Mac RAM | Guidance |
|---|---|
| 8 GB | Not practical for either variant |
Custom Fields
Special features defined by the model author
Enable Thinking
: boolean
(default=true)
Enables reasoning before the answer. Turn off for a direct answer.
Parameters
Custom configuration options included with this model
Sources
The underlying model files this model uses
| 16 GB |
| Estimated practical minimum: short text chats, small context, few other apps open |
| 24 GB+ | Recommended: more headroom for images, tools and longer conversations |
These are estimates, not verified minimums. Both variants were tested on a Mac with 64 GiB unified memory; 16 GB and 24 GB configurations were not tested. Actual memory use depends on context length, image resolution, temporary buffers and other running apps. Very long contexts can require substantially more memory. These recommendations do not guarantee that the full configured 262,144-token context will fit.
Compact saves about 187 MB in stored and loaded weights, which does not change the suggested RAM tier.
Use the Download Options selector to choose a variant. Both use native prism_hadamard_qwen35 schema 2 and require an MLX engine with its Hadamard transforms. The final Compact pack and catalog controls were tested with LM Studio 0.4.26+4 on Apple Silicon. Raw Hugging Face repository details in that app version offer one MLX choice and do not render the Thinking badge; this catalog supplies the grouped entry and explicit capability metadata.
Original and Compact each passed 20 targeted synthetic checks covering text, bounded thinking, tools with a mock round trip, colors, shapes, counting, OCR at five font sizes and a chart. LM Studio thinking on/off, a native tool call and round trip, and the public OCR fixture also passed. These checks are not a general quality benchmark, full-context validation or total RAM requirement. See the linked Hugging Face cards for reproducible evidence and limitations.
The language checkpoint is OS-Software's Uncensored Heretic v2 derivative of Prism ML's Bonsai 2 model. Credits are retained for OS-Software, Prism ML, Qwen/Alibaba Cloud, p-e-w's Heretic project and the Bonsai demo. Model redistribution uses Apache 2.0; see the full LICENSE and NOTICE files in each weight repository. The source publisher describes reduced safety alignment and recommends research and experimentation. The Compact pack does not change or restore language-model safety alignment.
The catalog contains public repository references and portable model controls. It contains no personal model paths, credentials, machine logs, MCP settings or local launch ports.
| 16 GB |
| Estimated practical minimum: short text chats, small context, few other apps open |
| 24 GB+ | Recommended: more headroom for images, tools and longer conversations |
These are estimates, not verified minimums. Both variants were tested on a Mac with 64 GiB unified memory; 16 GB and 24 GB configurations were not tested. Actual memory use depends on context length, image resolution, temporary buffers and other running apps. Very long contexts can require substantially more memory. These recommendations do not guarantee that the full configured 262,144-token context will fit.
Compact saves about 187 MB in stored and loaded weights, which does not change the suggested RAM tier.
Use the Download Options selector to choose a variant. Both use native prism_hadamard_qwen35 schema 2 and require an MLX engine with its Hadamard transforms. The final Compact pack and catalog controls were tested with LM Studio 0.4.26+4 on Apple Silicon. Raw Hugging Face repository details in that app version offer one MLX choice and do not render the Thinking badge; this catalog supplies the grouped entry and explicit capability metadata.
Original and Compact each passed 20 targeted synthetic checks covering text, bounded thinking, tools with a mock round trip, colors, shapes, counting, OCR at five font sizes and a chart. LM Studio thinking on/off, a native tool call and round trip, and the public OCR fixture also passed. These checks are not a general quality benchmark, full-context validation or total RAM requirement. See the linked Hugging Face cards for reproducible evidence and limitations.
The language checkpoint is OS-Software's Uncensored Heretic v2 derivative of Prism ML's Bonsai 2 model. Credits are retained for OS-Software, Prism ML, Qwen/Alibaba Cloud, p-e-w's Heretic project and the Bonsai demo. Model redistribution uses Apache 2.0; see the full LICENSE and NOTICE files in each weight repository. The source publisher describes reduced safety alignment and recommends research and experimentation. The Compact pack does not change or restore language-model safety alignment.
The catalog contains public repository references and portable model controls. It contains no personal model paths, credentials, machine logs, MCP settings or local launch ports.