GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
thanet-s
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
thanet-s
Model Tree
Base
deepreinforce-ai/Ornith-1.0-35B
Fine-tuned
this model
Input Modalities
Text
Image
Video
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
The Docker-contained Heretic project used to create this checkpoint is available on GitHub: thanet-s/Ornith-1.0-35B-heretic.
The Hugging Face model tree relation is set to finetune so this derivative is
linked under the base model's Finetunes section. The actual modification method
was Heretic abliteration rather than SFT.
| Parameter | Value |
|---|---|
| direction_scope | per layer |
| direction_index | 30.38 |
| attn.o_proj.max_weight | 1.49 |
| attn.o_proj.max_weight_position | 23.84 |
| attn.o_proj.min_weight | 1.20 |
| attn.o_proj.min_weight_distance | 23.07 |
These numbers are from the Heretic refusal probes used during the local run. They are not broad capability, safety, or coding benchmarks.
| Metric | This model | Original model (deepreinforce-ai/Ornith-1.0-35B) |
|---|---|---|
| KL divergence | 0.0063 | 0 (by definition) |
| Refusals | 53/100 | 90/100 |
| Item | Value |
|---|---|
| Heretic version | 1.4.0 |
| Source model | deepreinforce-ai/Ornith-1.0-35B |
| Export strategy | merge |
| Weights | BF16 safetensors |
| Quantization | none |
| Shards | 16 |
| Indexed tensors | 31,666 |
The exported checkpoint was structurally verified against its safetensors index and reloaded successfully with Transformers.
Ornith-1.0-35B is a reasoning model. Assistant turns may include
<think>...</think> content. When serving through an OpenAI-compatible runtime,
use a Qwen3-compatible reasoning parser if the runtime supports one.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "thanet-s/Ornith-1.0-35B-heretic"tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(model_id,dtype=torch.bfloat16,device_map="auto",trust_remote_code=True,)
If your Transformers build resolves this model family through the image-text-to-text auto class, use:
python
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessormodel_id = "thanet-s/Ornith-1.0-35B-heretic"processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForImageTextToText.from_pretrained(model_id,dtype=torch.bfloat16,device_map="auto",trust_remote_code=True,)
This checkpoint was modified to reduce refusals under Heretic's internal probes. It may be less likely to refuse requests than the base model and can produce harmful, incorrect, or policy-unsafe content. Evaluate carefully before any deployment.
No broad benchmark suite has been run for this checkpoint.
This derivative follows the base model license metadata, mit. Check the base
model card and upstream dependency terms before redistribution or deployment.