Qwen3-Coder-Next with Hermes Agent
FP8 on vLLM. Not tested yet
Task completion
Not tested yet
Tool-call success
Not tested yet
Decode tok/s
Not tested yet
Prefill tok/s
Not tested yet
Server memory
Not tested yet
Avg turns
Not tested yet
Setup
- Model
- Qwen3-Coder-Next
- Family / maker
- Qwen · Alibaba Qwen
- Parameters
- To be confirmed
- Architecture
- MoE
- Weights
- Exact repo confirmed at test time
- Quant
- FP8
- Engine
- vLLM (version recorded at test time)
- Hermes Agent
- Version recorded at test time
- Tested
- Not yet
- Context
- 64K (planned)
- Memory tier
- ≤128 GB (estimate)
Parameter count to be confirmed from the model card.
Per-task results
Results for each of the 10 benchmark tasks will be listed here once tested.
Copy-paste config
The exact vLLM launch command will be published once it has been verified in testing. We only publish configs we have run.
Point Hermes Agent at it
# Point Hermes Agent at a local OpenAI-compatible server
hermes config set model.provider custom
hermes config set model.base_url http://localhost:8000/v1
hermes config set model.default <served-model-id>
hermes -z "Reply exactly HERMES_OK"Generic Hermes Agent setup for any local OpenAI-compatible endpoint (from the Hermes Agent docs). Replace <served-model-id> with the id from curl -s http://localhost:8000/v1/models.
Other Qwen setups
- Qwen3.6-27B · Q6_K (GGUF) · llama.cpp
- Qwen3.6-35B-A3B · UD-Q6_K (GGUF) · llama.cpp
- Qwen3.5-122B-A10B · NVFP4 · vLLM
- Qwen3.8 35B-A3B · NVFP4 · vLLM