GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
hirundo-io
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
hirundo-io
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Standalone merged checkpoint derived from the Qwen team's
Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.
The existing RSCH-76 run qwen35-9b-g11-full-20261021-r2 checkpoint-360 LoRA was
already merged and verified before this upload. The run ended at step 480;
checkpoint 360 is the selected saved checkpoint. No training, merge, model
loading or evaluation was repeated for this upload.
17bb3139b82ea6b5af57bf02482a5b2ba49f974cc0e00c5c09b7e1d21b6cb848.85516b834d3ca4eb10087cf0f15834fb465d87d165ba53fbc77cf59c59e682a8.6667912fde32fe0d42773c0b55334274ec445ed1325a1f8fd30320c7748779d0.MERGED_SHA256SUMS file SHA256: 1288e9ca119432711beedefba427ae4697db7c131680d7e1e949353359065a1c.
This identifies the checksum-list bytes, not one weight file.The existing 9B results passed 182 independent CPU artifact-audit checks.
The audit used the saved evaluation records; it is not a general model safety
certification. RSCH-76 records completed MLflow evaluation upload/readback for
Base 23040eeb561147a699963fc61e34f306 and Westernized
ffaa0b4c81a9490b823d48c57aa5ca82. Training run
e97bdf278fc944ed85667c31f7aeb1f0 and parent
5f0761826bc54aa4b4536960e2d85d46 terminal FINISHED states were confirmed by the
owner in the UI; this upload does not claim a new independent MLflow API check.
See RSCH-76.
The tested scope was CCPC-500 and XSTest. The 9B native Inspect/Gemma/8192-token XSTest protocol was approved retrospectively for that session and is not directly comparable to GPT-4o/256-token results. Three Westernized CCPC responses were truncated; XSTest had zero truncations. These results do not establish general safety, safety equivalence, utility preservation or suitability for deployment.
Owner decision D-10: Thinking OFF (enable_thinking=false), BF16, vLLM reasoning
parser qwen3. These are explicit runtime settings; the inherited template does
not itself enforce Thinking OFF. No adapter is needed to load this merged model.
Qwen3.5-9B was developed by the Qwen team, Alibaba Cloud. The language-model weights were modified by the selected LoRA merge; inherited nonweight files are unchanged except this model card. Apache License 2.0 applies; see LICENSE.