GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Alex5000505
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Alex5000505
Model Tree
orcarouter/Qwen3.8-27B-Uncensored
Input Modalities
Output Modalities
Supported Functionality
orcarouter/Qwen3.8-27B-Uncensored @ 404ea47aaa5d8a8b00049c9e9750089aca011ab2.orcarouter/Qwen3.8-27B-Uncensored-FP8 @ 0f3cdb83820a8190ffedaef5b29cf4a635e49b4d.51de53e48ccae8804f8fe1198b7cf89475c5c4f4.65d4e2785addcec6bc8eca3af046006ec2725b6824e9cd5533170344808c675f.Requires vLLM support for Qwen3_5ForConditionalGeneration and ModelOpt MIXED_PRECISION. The deployed service used a patched vLLM build based on commit 360f33f61ce02bda04bcb7ea876f1d2e5af8d06d on SM120. Those runtime patches, endpoint credentials and TLS configuration are NOT model weights and are NOT bundled here. Exact deployed performance is not guaranteed with an arbitrary upstream installation.
Native MTP weights are present but must be enabled in the runtime. Vision weights are present; limits such as 32 images/request and KV-cache precision are server settings, not inherent limits of this checkpoint. The serialized head remains preserved even when a runtime applies an accelerated head path.
A bounded 24-task comparison against the previous FP8 checkpoint scored 21/24 vs 22/24 with no new failures in those tests. This is NOT proof of zero quality loss, broad benchmark superiority, or BF16 equivalence. The paired native BF16 comparison was interrupted to restore endpoint availability and has not established mixed-vs-BF16 equivalence.
SHA256SUMS verifies the uploaded model and metadata bytes. No conversations, request logs, API keys, TLS secrets, failed quantization variants or experimental retuned draft weights are included.
Start with reproduce/START-HERE-RU.md. Includes ready-weight download/verification, portable vLLM launch and live HTTP smoke, original 2048-record calibration IDs/config, portable BF16-to-NVFP4 gate/up calibration, checkpoint composer and preservation/corruption tests, and instructions for a receiving LLM agent.
The complete optimized custom runtime is not a portable container in this repository; RUNTIME-BOUNDARY.md states the remaining boundary. No claim of fresh-machine or TPS equivalence. See bundle validation.
The exact deployed source/native dependency overlay and external acceleration patches are now in runtime/START-HERE-RU.md, together with Dockerfile, weight verification, TLS/IP-allowlist launcher, checksums and validation boundaries. This is not a tested docker-save image: clean-image build and new-host inference/TPS qualification remain pending. No owner secrets, Bifrost code or private logs are included.