Instructions to use dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter

SGLang

How to use dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter with Docker Model Runner:
```
docker model run hf.co/dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter
```

🚀 J.A.R.V.I.S. TITAN 14.8B MoE — MILESTONE M2 (PHASE 2) TRI-BRID ADAPTER

Official Milestone M2 (Phase 2) weights for the J.A.R.V.I.S. Titan 14.8B DeepSeekMoE + Tri-Brid Memory Architecture, trained natively on Google Cloud TPU v5e-8 using Google Pallas TPU on-chip VMEM kernels.

🏛️ Tri-Brid Architecture Overview

J.A.R.V.I.S. Titan M2 integrates a 3-tier memory hierarchy across 7 strategic layers ([3, 7, 11, 15, 19, 23, 27]):

Tier 1: Sliding Window Attention (SWA) — Standard local attention ($W = 4096$) capturing immediate syntactic and discourse flow.
Tier 2: Salient Exact KV Reservoir ($K_{\text{res}} = 1024$) — Verbatim exact key-value buffer preserving past entities, UUIDs, code definitions, and needles.
Tier 3: Titans Neural Long-Term Memory ($d_{\text{mem}} = 3584$) — Associative fast-weight neural memory updating via test-time gradient momentum ($\mu = 0.95$, adaptive decay $lpha_t$).
MAG-3 (Memory-Augmented Gating) — Adaptive dynamic softmax gating routing forward representations: $$\mathbf{y}t = g{\text{local}} \cdot \text{Attn}(\mathbf{x}t) + g{\text{res}} \cdot \text{Reservoir}(\mathbf{x}t) + g{\text{mem}} \cdot \text{Titans}(\mathbf{x}_t)$$

📊 Evaluation Matrix (Milestone M2 Certified)

Evaluation Track	Context	Accuracy / Score	Perplexity ($PPL$)	Base MoE Baseline
Track 1: Olympiad Math Proofs	1K	100.0% Passed	21.39	40.5
Track 1: Agentic Tool Loops	2K	94.1% Passed	21.65	81.4
Track 1: DeepSeek-R1 Long CoT	4K	100.0% Passed	22.17	345.1
Track 1: Neural Codebase Architecture	8K	62.7% Passed	23.20	6,868.0
Track 1: Repo-Scale RoPE Tracing	16K	100.0% Passed	25.32	3.3M (Collapsed)
Track 1: Scientific Paper Synthesis	32K	100.0% Passed	29.99	485M (Collapsed)
Track 2: Exact Verbatim Retrieval	4K–32K	100.0% Recall	—	0.0%
Track 3: Multi-Needle Distractor Recall	4K–32K	100.0% Recall	—	0.0%

📦 Model Files

jarvis_titan_m2_tribrid_adapter.safetensors: 90.04M parameters across 7 layers (171.76 MB, 105 tensors).
adapter_config.json: Hardware-native hyperparameters and layer routing mappings.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for dhanesh-hf/Jarvis-Titan-M2-TriBrid-Adapter

Base model

dhanesh-hf/Jarvis-Titan-V14-MoE-Merged

Finetuned

(3)

this model