Sage-A1.5-MXFP4_MOE-MTP-GGUF

Sage-A1.5 is a 35B Mixture-of-Experts (MoE) multimodal agentic model from Shanghai AI Laboratory designed for long-horizon scientific problem solving, complex engineering workflows, and open scientific discovery. Building upon Agents-A1, it pairs native tool integration with two opt-in execution paradigms: Agentic Orchestration, which dispatches parallel sub-agents with isolated contexts (spawn_subagent) to decompose and investigate independent sub-questions concurrently, and Verification Mode, an in-trajectory self-review mechanism that audits evidence and revises unproductive strategies directly inside the model's chain of thought without an external critic. Equipped with a native 256K context window (extensible to 512K via YaRN) and Multi-Token Prediction (MTP) support for speculative decoding in vLLM and SGLang, Sage-A1.5 rivals frontier-scale systems despite its 35B footprint—achieving overall SOTA on GAIA (95.1), OmniDocBench (91.5), XBench-DS-2510 (92.0), PinchBench (90.0), IFBench (81.6), and FrontierScience-Olympiad (81.0), while leading comparable 35B models on SWE-bench Pro (61.2) and BrowseComp (82.8).

MTP — Speculative Decoding: Yes

Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.

Model Files

File Name	Quant Type	File Size	File Link	Description
Sage-A1.5.BF16.gguf	BF16	71.1 GB	Link	Full BF16 weights. Highest quality, largest file size.
Sage-A1.5.MXFP4_MOE.gguf	MXFP4_MOE	20.3 GB	Link	Mixed precision / MoE optimized quantization, highly efficient size.
Sage-A1.5.Q3_K_L.gguf	Q3_K_L	18.6 GB	Link	Lower quality but usable, good for low RAM availability.
Sage-A1.5.Q5_K_M.gguf	Q5_K_M	25.3 GB	Link	High quality, recommended.
Sage-A1.5.Q6_K.gguf	Q6_K	29.2 GB	Link	Very high quality, near perfect, recommended.
Sage-A1.5.mmproj-bf16.gguf	mmproj-bf16	903 MB	Link	Multimodal projection file in BF16 format. Used for vision/language models.

MXFP4-MOE Test Run

Tested MXFP4_MOE on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading.

Performance Metrics

Metric	Value
Prompt Processing Speed	94.1 tokens/s
Generation Speed	270.4 tokens/s
Total Wall Time	20.216 seconds
Speculative Decoding	draft-mtp
Maximum Draft Tokens	2
Context Length	8,192 tokens
KV Cache (K)	q8_0
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU VRAM Capacity	96 GB
Peak GPU VRAM Usage	19.72 GiB
Peak GPU Utilization	81%
GPU Layers Offloaded	99 (-ngl 99)
Random Seed	42

Run Inference

./build/bin/llama-cli \
  -m /path/to/models/Sage-A1.5.MXFP4_MOE.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  -ngl 99 \
  -c 8192 \
  --cache-type-k q8_0 \
  --seed 42 \
  -n 4096 \
  -st \
  -p "Your prompt here"

Notes

• Full GPU offloading: -ngl 99 requests offloading up to 99 model layers to the GPU. Actual offloading depends on the model architecture and the number of layers available.
• Speculative decoding: --spec-type draft-mtp enables MTP-based speculative decoding, with a maximum of 2 draft tokens per step.
• KV cache: --cache-type-k q8_0 configures the K cache to use the q8_0 quantization format.
• Context length: -c 8192 sets the context size to 8,192 tokens.
• Generation length: -n 4096 sets the maximum number of tokens to generate.
• Random seed: --seed 42 sets the random seed for reproducibility.

Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

---

Quantized using llama.cpp by prithivMLmods