Spark-Hermes-3.8-27B

Spark-Hermes-3.8-27B

Hermes × Qwen3.8-27B co-trained agent model from Gittensor Model Hub.

This repo is the model card first. Weights are not uploaded yet. They will land here.

Built to pair with SparkInfer on RTX 5090 and RTX PRO 6000: few SOTA models, optimized all the way, already 2–3× llama.cpp.

---

What this is

Qwen3.8-27B is a dense 27B vision-language model (hybrid Gated DeltaNet + Gated Attention, native 262,144 context). Hermes is the agent post-train lineage: function calling, structured output, multi-turn control.

Spark-Hermes-3.8-27B co-trains those two: Qwen3.8-27B capability with Hermes agent behavior — for tools, JSON, and long-horizon work on a single workstation GPU.

Sister checkpoint on this hub (already shipped): Qwen3.8-27B-NVFP4-RTX5090 — NVIDIA ModelOpt NVFP4 of the Qwen3.8-27B base for RTX 5090 (full 256k on 32 GB).

---

Status

Weights are not in this repository yet. This page is the card so the name, lineage, and intended use are public while training / export finishes.

When weights land, this README will add serve commands (SparkInfer first, then vLLM) and eval.

---

Intended use

• Agentic assistants (tools, function calling, structured output)
• Long-running research / coding / worker loops
• Local serve on RTX 5090 and RTX PRO 6000 via SparkInfer

Not a drop-in replacement for the NVFP4 base checkpoint until weights are published here.

---

Lineage

Qwen/Qwen3.8-27B  ×  Hermes (agent co-train)
        └── Spark-Hermes-3.8-27B   ← this repo (card first)

Related: gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090

---

License

Apache 2.0, same as Qwen/Qwen3.8-27B.