Marketplace · Text & chat
granite3.1-moe:1b
Call granite3.1-moe:1b through one OpenAI-compatible API. It's served by 1 host on the Inpherio grid, each at a price it sets — your request goes to the cheapest healthy one, and fails over to the next if it's down.
At a glance.
tools
from openai import OpenAI
client = OpenAI(base_url="https://inpherio.co.uk/v1",
api_key=INPHERIO_API_KEY)
reply = client.chat.completions.create(
model="granite3.1-moe:1b",
messages=[{"role": "user", "content": "Hello"}],
)granite3.1-moe:1b in detail.
The IBM Granite 1B and 3B models are long-context mixture of experts (MoE) Granite models from IBM designed for low latency usage.
- Published by
- mesmac
- Parameters
- 1.8B
- Download size
- 1.4 GB (as packaged for Ollama)
- Quantisation
- Q8_0
- Format
- gguf
- Architecture
- granitemoe
- Licence
- Apache 2.0
- Based on
- ibm-granite/granite-3.1-1b-a400m-instruct
- Library
- pytorch
- First published
- 26 Oct 2026
- Last updated
- 26 Oct 2026
- Downloads (30 days)
- 71 on Hugging Face
mcl4-coe ncm next-concept-modeling mixture-of-experts cognitive-operator-ensemble markov-local-attention granite nvfp4
The model card
Published by the model's authors on Hugging Face — shown as they wrote it, unedited by Inpherio.
MCL4-COE-v7-Granite-3.1-1B-NCM-MoE (Step 9,000, 1.77B Parameters)
Code & Technical Documentation: https://github.com/mESmaC/mcl4-coe-ncm
This checkpoint grafts a 6-level Donor-Anchored Multiscale Residual VQ-VAE (NCM, $\ell_k \in [1, 4, 11, 24, 41, 64]$, $V_k = 1024$), Scale-Adaptive MoE Routing, Markov Chain Local-4 (MCL4) sparse graph attention (markov_local_tokens=8 anchors + 16 Markov transitions = 24 attended positions), and the 9-Operator Single-Softmax Cognitive Operator Ensemble (COE) onto ibm-granite/granite-3.1-1b-a400m-instruct (12 encoder + 12 decoder layers, d_model=1024, n_heads=16, vocab_size=49155).
Architecture & Training Trajectory (Step 5,000 -> Step 9,000)
- Phase 1 (
Step 5,000 -> 5,500,E=32, top_k=8Frozen-Donor NCM Graft Validation): - Scale-Normalized Dead-Code Restart & Codebook-RMS Query Normalization eliminated the $20\times$ Laplace-smoothed EMA codebook norm inflation (
||E||_2: 0.7 -> 19.6) across multiscale levels $\ell_1 \dots \ell_K$ and dropped initial Step-1_TiedCodebookHeadgenerator cross-entropy from51.44to7.59. - Achieved
62.92%active codebook utilization across all $6 \times 1024$ residual codes and84.4%mean /89.5%peak exact token reconstruction at the finest scale ($t_5 = 64$), with quantization misses decoding to nearest semantic synonyms in Granite 3.1's frozen embedding space. - Reduced
decision_lossby26.4%(0.6567 -> 0.4831) and raisedOrnith-2.5verification pass rates from71.6%to80.5%. - Phase 2 (
Step 5,500 -> 9,000, Mitotic ExpansionE=32 -> 40, top_k=8NVFP4): - Mitotically expanded to
E=40experts (1,773,191,615parameters =1.77B) with 8 mitotic twin pairs ([13,32], [31,33], [19,34], [8,35], [30,36], [24,37], [28,38], [0,39]) across27.55Mcumulative tokens and46,723,200fused multi-expert NVFP4 GEMMs. - Step 9,000 Final Telemetry (
config.json): ncm_codebook_utilization:77.78%(up from62.92%at Step 5,500)ncm_recon_accuracy:85.35%(ncm_vq_recon_loss = 0.6523)ornith_pass_rate:99.66%(ornith_rollout_reward = 0.9971,ornith_total_pp = 10,606,273.6,236acceptedDream-RSIrevisions across8,574,600dreamed replays)- Held-Out Validation (
val_bpb):0.7535(val_bpb_typed_decision = 0.4981,val_bpb_uta_trace = 0.7763,val_bpb_pretrain_lm = 1.0007)
Quickstart (Safetensors)
pip install safetensors huggingface_hub "git+https://github.com/mESmaC/mcl4-coe-ncm.git"
import torch
from mcl4_coe import load_pretrained_mcl4_coe_ncm
model = load_pretrained_mcl4_coe_ncm(
"mesmac/MCL4-COE-v7-Granite-3.1-1B-NCM-MoE",
device="cuda" if torch.cuda.is_available() else "cpu",
dtype="bfloat16",
)
print("Loaded MoE experts:", model.backbone.config.num_experts, "| NCM levels:", model.ncm_system.levels)Sources: mesmac/MCL4-COE-v7-Granite-3.1-1B-NCM-MoE on Hugging Face · granite3.1-moe:1b on Ollama. Collected from public sources on 14 Nov 2026; the host serving it may run a different build or quantisation — the marketplace shows each host's own measurements.
Who's serving it.
Every host benchmarked this model on its own hardware before serving it. Prices are set by each host; health is scored by the platform from what it observes.
| Host | Tier | Price | Measured speed | Health | State | Rent |
|---|---|---|---|---|---|---|
| GTX 1080 Ti (11G) | L1 | £0.03 in · £0.06 out | 55 tok/s · 23 ms first token | 78 | loads on request | £0.60/h |
Live data, refreshed on every visit. How billing works.
Room on the grid for your machine.
Right now one machine serves granite3.1-moe:1b. If it drops, so does every request for it — a second host is what turns it into failover, and the grid needs that more than another copy. The router weighs price, measured speed, health and whether the model is already loaded — so a host with a sharper price or quicker hardware wins the next call.
Set your own price — you keep 85% of what it earns, and new hosts start with 12 months fee-free (up to £250 of earnings). A payout fee of £2 + 1% per payout to a UK bank account, £2 + 2% to any other country applies when you're paid. Install the connector, load granite3.1-moe:1b, and the platform benchmarks it on your hardware before your first request.