IBM · Granite 4.0

Granite 4.0 H Tiny

Verified by Badgr · Serve

Granite 4.0 H Tiny is a 6.9B chat and text-generation model available through supported Badgr execution routes.

Last validated or meaningfully updated: 2026-10-09

ChatLong contextFine-tunable
Serve Granite 4.0 H Tiny nowGet help running this

✓ Verified by Badgr · SGLang · Serve

Last verified 2026-10-09

badgr serve --image lmsysorg/sglang:nightly-dev-20261005-f70e8c68 --gpu auto --port 30000 --health-path /health --max-cost 2 --startup-timeout 40 --cmd 'python3 -m sglang.launch_server --model-path ibm-granite/granite-4.0-h-tiny --host 0.0.0.0 --port 30000 --enable-deterministic-inference'
GPU
NVIDIA GeForce RTX 3090Reported by the provider, not confirmed inside the container
Time to verified
414.9s
Deployment
dep-6d3e5d79c9
  • ✓ SGLang server started
  • ✓ /health returned HTTP 200
  • ✓ A repetition_penalty completion request returned text
  • ✓ final_state = VERIFIED
  • ✓ Teardown completed

Served with --enable-deterministic-inference. A completion request with repetition_penalty 1.06 (the trigger in sgl-project/sglang#43061) returned normally and the server stayed healthy on this image.

GitHub issue this run was for: sgl-project/sglang#43061

Raw evidence
final_state=VERIFIED failure_class=None providers_tried=1 spend_usd=0.0329 elapsed=414.9s

Run from Badgr's local development environment. Documents that this deployment path works end-to-end for this model.

Availability

Not yet Badgr AI API

Not currently validated

✓ Dedicated endpoint

Available through Badgr

✓ Custom GPU deployment

Available through Badgr

Deployment profile

Architecture
GraniteMoeHybridForCausalLM
Licence
Apache-2.0
Minimum VRAM
~18GB
Context
128K tokens

Recommended: RTX 4090 24GB. VRAM is an estimate and increases with context, cache, concurrency, and runtime overhead.

Model identity

Updated
11 months ago
HuggingFace revision
791e0d3d28

Popularity and trust

Downloads (last month)

52.4K

Likes

210

Spaces using this model

5

Files and formats

Weight formats

Safetensors

Repository files

15

Tokenizer

BPE

Chat template

Available

GPU deployment scenarios

Estimated from parameter count and quantisation. Will switch to “Verified by Badgr” once a real run backs a tier.

Testing

RTX 4090 24GB

Short context, low concurrency

Small production

L40S 48GB

8K context, 1-4 concurrent requests

Higher throughput

A100 80GB

32K context, continuous batching

Large production

RTX 4090 24GB

Higher concurrency

Estimated — multimodal inputs, long context, and concurrency all increase real VRAM use beyond this estimate.

Also runs well on L40S 48GB, A100 40GB. Available in United States, Europe.

Serving compatibility

RuntimeStatusNotes
vLLMEstimatedUsed for Badgr dedicated endpoints
TransformersDeclared by sourceBasic fallback
llama.cppRequires GGUF conversionNo GGUF weights found in repo
SGLangVerified by BadgrConfirmed by a real badgr serve run (deployment dep-6d3e5d79c9)
OllamaUnknownNot evaluated
DiffusersUnknownNot evaluated
PyTorchUnknownNot evaluated
TensorRT-LLMUnknownConversion may be required
TGIUnknownNot evaluated

Ready-to-run Badgr configurations

Quick test

badgr serve ibm-granite/granite-4.0-h-tiny --max-cost 2

Validated dedicated endpoint command

badgr serve ibm-granite/granite-4.0-h-tiny --gpu RTX 4090 --max-cost 10

Advanced configuration (defaults to automatic)

  • Runtime
  • Quantisation
  • GPU and GPU count
  • Context length
  • Maximum concurrency
  • Region
  • Maximum hourly spend
  • Persistent or capped runtime

Estimated pricing

Provider estimate: cold start

2–5 min

Provider estimate: endpoint cost

$0.18 – $0.45/hr

Idle cost

$0 when stopped

Per-token throughput cost is not shown here: Badgr has not benchmarked this model yet, and this page does not display figures it cannot back with real data.

Recommended Badgr routes

RTX 4090 24GB · United States

Available

from $0.18/hr

24GB VRAM/GPU16GB RAM1440.9GB storage0.02Gbps networkCUDA 13.2

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: just now

Check live route
View details →

RTX 4090 24GB · United States

Available

from $0.43/hr

2x GPU24GB VRAM/GPUSame host8 vCPU31.3GB RAM704GB storage0.81Gbps networkCUDA 13.2

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: just now

Check live route
View details →

RTX 4090 24GB · United States

Available

from $0.44/hr

24GB VRAM/GPU

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: just now

Check live route
View details →

RTX 4090 24GB · Europe

Estimated

starting from $0.45/hr

Badgr estimated starting price

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: Live marketplace price not available yet

Check live route
View details →

RTX 4090 24GB · Europe

Estimated

starting from $0.45/hr

Badgr estimated starting price

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: Live marketplace price not available yet

Check live route
View details →

RTX 4090 24GB · Europe

Estimated

starting from $0.45/hr

Badgr estimated starting price

Best for

ComfyUI, Batch Inference, Qwen 2.5 7B Instruct

Startup: 2–5 min

Reliability: Standard

Last checked: Live marketplace price not available yet

Check live route
View details →

Known limitations

  • Memory use increases with context length and concurrency.

Related troubleshooting