GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
AlexWortega
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AlexWortega
Model Tree
Base
Qwen/Qwen3.5-4B
Fine-tuned
this model
Input Modalities
Text
Image
Video
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
bash
python -m sglang.launch_server \--model-path AlexWortega/qwen35-4b-soyuz-abliterated-v3-multi \--dtype bfloat16 --trust-remote-code \--tool-call-parser hermes \--chat-template hermes_qwen.jinja
(hermes parser is needed for the <tool_call>{...}</tool_call> → OpenAI tool_calls conversion — without it agent benches see zero tool calls.)
reward=1.0) + 60 cleaned FAIL trajectories from soyuz's own evals (claw-eval, tbench-2, MMLU-Pi-agent). Fail trajectories filtered by Gemini-3-flash to keep only CLEAN_FAIL labels (235 of 246 negatives).Qwen3_5ForCausalLM).mean(refuse) - mean(comply), normalise; pick best layer via AUC.o_proj.weight and down_proj.weight columns) against the direction, optionally blended with strength α: W ← W − α · (W − W_orth).Qwen3_5ForConditionalGeneration arch so sglang can serve them (vision tower preserved from base; only language_model.* weights are abliterated).| Variant | tbench-17 | HA20 | Card |
|---|---|---|---|
baseline qwen35-4b-soyuz (LoRA) | 5/17 | 4/20 | link |
qwen35-4b-soyuz-abliterated-v2 (single-L, s=0.5) | 3/17 | 8/20 | link |
qwen35-4b-soyuz-abliterated-v3-multi (per-layer, s=0.5) |
v2 = highest HA20 (2× baseline). v3 picks up disjoint HA20 tasks (HA-01/02 memory-specific) that v2 misses.
W&B + raw eval logs: https://wandb.ai/alexwortega/vae-llm-agents (training base).
| 2/17 |
| 6/20 |
| link |