Downloads · 30 days
15
3% of all-time downloads
eddy1111111/Qwen3.5-122B-A10B-ReActsft-nvfp4
Qwen3.5-122B-A10B-ReActsft-nvfp4 is a text generation model from eddy1111111. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as other.
A reinforcement-trained Qwen3.5-122B-A10B with enhanced browser agent interleaved reasoning capabilities, quantized to NVFP4 for efficient deployment.
Downloads · 30 days
15
3% of all-time downloads
All-time downloads
465
Public
Parameters
62.3B
78.5 GB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors78.4 GB · 100%
How the weights are stored.
U858B · 81%
From the Hugging Face model README
A reinforcement-trained Qwen3.5-122B-A10B with enhanced browser agent interleaved reasoning capabilities, quantized to NVFP4 for efficient deployment.
| Attribute | Value |
|---|---|
| Base Model | Qwen3.5-122B-A10B (MoE, 122B total / 10B active) |
| Training Method | Reinforcement Learning (browser agent focus) |
| Quantization | NVFP4 (NVIDIA TensorRT-Model-Optimizer, custom-patched) |
| Training Data | <500 fully human-crafted samples |
| Training Time | ~5 hours |
| Training Hardware | NVIDIA RTX PRO 6000 (Blackwell, 96 GB) |
| Context Length | 262,144 tokens |
| MTP Layers | Excluded (acceptance rate too low post-quantization) |
This model is reinforcement-trained over the base Qwen3.5-122B-A10B, with the primary focus on browser agent interleaved reasoning. The reinforcement is not a general-purpose enhancement — it specifically targets the model's ability to operate as a browser automation agent through multi-step thinking-action loops.
The base Qwen3.5-122B-A10B can call browser tools, but it does so in a shallow, rapid-fire manner — typically 5 steps with no intermediate reasoning, just blind tool calls. After reinforcement training, this model can sustain 20-30 consecutive browser action rounds in a single session, with reflective thinking between each step.
The key improvements:
This model is not broadly reinforced across all task types. General reasoning, coding, math, and non-browser tool-calling capabilities remain at base model level. The training data and reward signal are specifically designed around browser agent sessions.
The training data consists of fewer than 500 entirely human-crafted samples (not synthetic, not auto-generated — every single sample was manually created from real browser agent sessions). Each sample follows the interleaved pattern:
thinking → action(browser_tool) → observation → thinking → action(browser_tool) → ... → thinking → final_response
Single-GPU deployment on any 96 GB card (RTX PRO 6000, A100-80GB with NVLink, etc.):
python -m sglang.launch_server \
--model-path /path/to /Qwen3.5-122B-A10B-ReActsft-nvfp4 \
--tp 1 \
--context-length 262144 \
--host 0.0.0.0 --port 8000
Measured throughput: ~83 tok/s (single RTX PRO 6000, FP4, TP1).
vLLM does not natively support the NVFP4 LoRA merged weight format out of the box. You will need to patch the model loader to handle the quantized LoRA-merged checkpoints. This is not a drop-in replacement — expect to modify vllm/model_executor/ for weight deserialization.
Same as the base model: Qwen License.