GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
lucataco
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
lucataco
Model Tree
deepreinforce-ai/Ornith-1.0-35B
Input Modalities
Output Modalities
Supported Functionality
5df2ed3f675c7beaa490328cc70bb573b65fb660transformers + fouroversixnvfp4static_6bfloat16lm_head plus 154 shape-incompatible linear modules are
kept in the original BF16 format and listed in modules_to_not_convert.The checkpoint was converted by streaming the original safetensor shards into
FourOverSix-compatible quantized tensors, so it can be loaded with recent
transformers builds that include FourOverSixConfig support and the
fouroversix package installed.
This checkpoint was smoke-tested with transformers==5.12.1,
torch==2.10.0+cu130, and fouroversix==1.0.5. The local FourOverSix install
also included the Qwen3.5 MoE compatibility fix for DataType property access
inside fouroversix/model/modules/qwen.py. If loading raises an error such as
DataType object is not callable, a missing scale_dtype, or missing
get_weight_scale_rule / get_activation_scale_rule, update FourOverSix to a
release containing that fix or patch that Qwen module.
Tested on an NVIDIA GB10 with transformers==5.12.1,
torch==2.10.0+cu130, and the patched fouroversix==1.0.5 runtime. The test
used batch size 1, greedy manual decoding, synthetic repeated-token context,
1024-token chunked prefill, and logits_to_keep=1.
| Context tokens | Decode speed |
|---|---|
| 1,024 | 8.95 tok/s |
| 8,192 | 9.08 tok/s |
| 32,768 | 6.64 tok/s |
| 100,000 | 7.06 tok/s |
The 100k context run completed, but used about 120.6 GiB of the GB10's 121.7 GiB visible CUDA memory, leaving roughly 1.1 GiB free.