MainStack/marvy-2-35B-MoE-GGUF

marvy-2-35B-MoE-GGUF

GGUF quants of marvy-2, a 35B-A3B Mixture-of-Experts model fine-tuned for the ServiceNow delivery lifecycle. ~3B active parameters per token; runs on a single consumer GPU or fast Apple Silicon thanks to the MoE sparsity.

GGUF quantizations for use with llama.cpp, Ollama, LM Studio, and compatible runtimes.

Released under Apache-2.0. Built with Qwen3.5 (Apache-2.0) via unsloth/Qwen3.6-35B-A3B and the Opus-distilled stamsam/...MTP base.

Files

Architecture notes

This is a hybrid SSM + MoE Transformer:

• —40 layers, mixed SSM (Mamba-style) and grouped-query attention blocks
• —256 routed experts per MoE layer, 3B active per token (A3B)
• —Shared expert per layer for common pathways
• —qwen3_5_moe architecture in llama.cpp; needs a recent build (see "Supported runtimes" below)

The Multi-Token Prediction (MTP) head present in the base model was not included in this GGUF — these quants are text-only causal LM. The base model's MoE expert weights and SSM blocks are preserved.

Quick start

Ollama

ollama run hf.co/MainStack/marvy-2-35B-MoE-GGUF:Q4_K_M

llama.cpp

llama-cli -hf MainStack/marvy-2-35B-MoE-GGUF:Q4_K_M \
  -p "Write a ServiceNow user story with acceptance criteria for P1 SLA escalation." \
  --temp 0.4 \
  -c 4096

LM Studio

Search the model catalog for marvy-2-35B-MoE-GGUF and download, or load the local .gguf file via "Open Model in Folder". LM Studio's OpenAI-compatible server is at http://localhost:1234/v1 by default.

Supported runtimes

The qwen3_5_moe architecture (with mixed SSM layers and Multi-Token Prediction in the base) is new. Verify your runtime supports it:

• —llama.cpp: master branch as of mid-2026 (commits containing Qwen3_5MoeForConditionalGeneration registration in conversion/qwen.py). The Homebrew formula may lag — clone upstream if you hit "unknown architecture" errors.
• —Ollama: ships its own llama.cpp; check that your Ollama version is recent.
• —LM Studio: uses its own bundled runtime; recent versions support Qwen3.5 MoE.

Trained on

• —v1 corpus: ServiceNow delivery lifecycle artifacts (SOW, SDD, stories, acceptance criteria, value hypothesis, ...) — same data marvy-1 used.
• —v2 corpus: extended to capability-to-epic mapping, mermaid diagram authoring, deployment package modeling, stakeholder mapping, story-to-UAT, and more. See EVAL.md in the repo root for per-task perplexity.

How this GGUF was built

LoRA adapter (rank 32, 350 steps, attention-only Q/K/V/O)
    +
bf16 base   (unsloth/Qwen3.6-35B-A3B)
    │   mlx_lm fuse
merged-bf16/  (14 shards, 65 GB safetensors with mlx_lm switch_mlp naming)
    │   scripts/marvy-v2-rename-moe-tensors.py  (bridge to HF-canonical names)
merged-bf16-hf/  (15 shards, 65 GB; switch_mlp → experts.gate_up_proj packed)
    │   llama.cpp/convert_hf_to_gguf.py --no-mtp --outtype f16
marvy-2-35B-MoE-F16.gguf  (65 GB, 733 tensors)
    │   llama-quantize
marvy-2-35B-MoE-Q4_K_M.gguf  (20 GB, 4.6 BPW)
marvy-2-35B-MoE-Q8_0.gguf    (34 GB, 8.52 BPW)

End-to-end build script: scripts/marvy-v2-35B-MoE-build-gguf.sh (in the source repo). The switch_mlp → experts rename is necessary because mlx_lm fuse and llama.cpp's converter use different naming conventions for routed MoE tensors.

License

Apache-2.0 (inherits from Qwen). See LICENSE and NOTICE in the repo root.

Citation

@misc{marvy-2-35B-MoE,
  title  = {marvy-2: A ServiceNow delivery-lifecycle MoE LLM},
  author = {MainStack},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/MainStack/marvy-2-35B-MoE-GGUF}}
}