GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
olegsmirnov
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
olegsmirnov
Model Tree
caiovicentino1/Eikos-27B
Input Modalities
Output Modalities
Supported Functionality
The pipeline is Eikos-27B's: same code, prompts, calibration and API. letter_adapter.py sees the 588-row head and
loads it (trimmed_head.py).
bash
hf download olegsmirnov/Eikos-27B-BuzzCut --local-dir Eikos-27B-BuzzCutpython Eikos-27B-BuzzCut/serve.py --model Eikos-27B-BuzzCut --port 8000python
import sys sys.path.insert(0, "Eikos-27B-BuzzCut")from decision_core import options_offrom letter_adapter import LetterAdapter ad = LetterAdapter( "Eikos-27B-BuzzCut", device="cuda", calib="Eikos-27B-BuzzCut/calib.json")state = "Your friend texts you \"I'm five minutes away\", but they're a zoomer."q = { "type": "choice", "instructions": "When will they arrive?", "criteria": { "on_the_dot": "In 5 minutes, on the dot", "fashionably_late": "In 7-10 minutes, fashionably late", "tomorrow": "Maybe tomorrow", "behind_you": "They're already there, right behind you", },}print(ad.dist(state, q, options_of(q))[0])# {'on_the_dot': 0.0645, 'fashionably_late': 0.4203, 'tomorrow': 0.3948, 'behind_you': 0.1204}The PyTorch backend of the release: serve.py, local_demo.py, LetterAdapter. In the transformers config one
vocab_size sizes both the input embeddings and lm_head, so a smaller head cannot be declared there; the adapter
loads the checkpoint with the size mismatch ignored and attaches the 588 rows, label j at column j
(trimmed_head.py), and a plain AutoModelForCausalLM.from_pretrained refuses the file. vLLM and SGLang
(serve_vllm.sh) build the full head from the config: use Eikos-27B there. The verify mode, mlx_decide.py and
images are not supported.
MIT (Eikos, by Caio Vicentino); Apache-2.0 for the Qwen3.8-27B base (LICENSE-Qwen, NOTICE).
import sys sys.path.insert(0, "Eikos-27B-BuzzCut")from decision_core import options_offrom letter_adapter import LetterAdapter ad = LetterAdapter( "Eikos-27B-BuzzCut", device="cuda", calib="Eikos-27B-BuzzCut/calib.json")state = "Your friend texts you \"I'm five minutes away\", but they're a zoomer."q = { "type": "choice", "instructions": "When will they arrive?", "criteria": { "on_the_dot": "In 5 minutes, on the dot", "fashionably_late": "In 7-10 minutes, fashionably late", "tomorrow": "Maybe tomorrow", "behind_you": "They're already there, right behind you", },}print(ad.dist(state, q, options_of(q))[0])# {'on_the_dot': 0.0645, 'fashionably_late': 0.4203, 'tomorrow': 0.3948, 'behind_you': 0.1204}