NSF National Deep Inference Fabric
Operated by Northeastern University with NCSA/UIUC
No-cost research and development inference across 114 models; not unrestricted production use.
Models mentioned
allenai/Olmo-3-1125-32BEleutherAI/gpt-j-6bgoogle/gemma-2-2bgoogle/gemma-2-2b-itgoogle/gemma-2-9bgoogle/gemma-2-9b-itmeta-llama/Llama-3.1-70Bmeta-llama/Llama-3.1-70B-Instructmeta-llama/Llama-3.1-405B-Instructmeta-llama/Llama-3.1-8Bmeta-llama/Llama-3.1-8B-Instructmeta-llama/Llama-3.2-1Bmeta-llama/Llama-3.2-1B-Instructmeta-llama/Llama-3.2-3Bmeta-llama/Llama-3.2-3B-Instructmeta-llama/Llama-3.3-70B-Instructopenai-community/gpt2Qwen/Qwen3-VL-8B-InstructQwen/Qwen2.5-7BQwen/Qwen2.5-7B-InstructQwen/Qwen3-4BQwen/Qwen2.5-Coder-7B-Instructmicrosoft/Phi-3.5-mini-instructmistralai/Mistral-Small-24B-Instruct-2501meta-llama/Llama-3.1-405Bnvidia/Qwen3.6-35B-A3B-NVFP4google/gemma-4-26B-A4BQwen/QwQ-32BQwen/Qwen2.5-14B-InstructQwen/Qwen3-32Bfacebook/esm2_t12_35M_UR50DQwen/Qwen3-4B-Thinking-2507Qwen/Qwen2.5-0.5Bibm-granite/granite-4.2-30bQwen/Qwen2.5-72Byen-av/olmo-3-7b-butterfly-refusalgoogle-bert/bert-large-uncasedkernels-community/triton_kernelstiiuae/Falcon3-7B-Baseopenai-community/gpt2-xllewtun/talkie-1930-13b-it-hfallenai/OLMo-7Bllava-hf/llava-interleave-qwen-0.5b-hfMaykeye/TinyLLama-v0Qwen/Qwen3.5-9Bopenai/gpt-oss-120bopenai/gpt-oss-20bzai-org/GLM-4.5-Airmistralai/Mistral-7B-Instruct-v0.3google/gemma-4-31Bneuronpedia/jacobian-lenshumain-ai/ALLaM-7B-Instruct-previewmeta-llama/Llama-2-7b-chat-hfEleutherAI/pythia-2.8bmeta-llama/Llama-2-7b-hfQwen/Qwen2-72B-Instructgoogle/gemma-4-12B-itmistralai/Mixtral-8x22B-v0.1Qwen/Qwen2.5-0.5B-Instructnvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16google/gemma-2-27bmeta-llama/Llama-2-7bdeepseek-ai/DeepSeek-V4-Flashgoogle/gemma-3-4b-ptgoogle/gemma-3-4b-itmosaicml/mpt-30bgoogle/gemma-3-27b-ptgoogle/gemma-3-12b-ptgoogle/gemma-3-1b-ptgoogle/gemma-3-12b-itfacebook/esmfold_v1microsoft/DialoGPT-smallgoogle/gemma-3-27b-itallenai/OLMo-2-0425-1Ballenai/Olmo-3.1-32B-ThinkQwen/Qwen3-8Bstabilityai/stable-diffusion-xl-base-1.0microsoft/DialoGPT-mediumstable-diffusion-v1-5/stable-diffusion-v1-5allenai/Olmo-3-7B-InstructShowing 80 of 114. Use the provider discovery URL for the complete live catalog.
Equivalent paid value
How this was valued: The published allowance or a defensible paid comparison is not precise enough to calculate.
Research access has no public numeric user allocation or commercial paid counterpart.
Limits and terms
- Published Numeric Limits
- No
What happens to your prompts?
No endpoint-wide public prompt-retention, training, operator-access, or deletion policy was found.
- Caveat
- NDIF documentation and project pages explain research access and instrumentation, but no public service terms or privacy policy defining prompt retention, research use, operator access, or deletion was found. Research workloads should be assumed observable unless a project agreement says otherwise.
No provider terms or privacy-policy link is captured in this snapshot. Review the provider’s current legal documents before sending sensitive data.
Eligibility
- Account Required
- Yes
- Payment Method Required
- No
- Allowed Use
- research
Primary sources
Before you build with NSF National Deep Inference Fabric
Read the classification narrowly
This record has current first-party evidence for some zero-cost hosted inference. The label describes the bounded offer supported by evidence on 2026-09-19; it does not guarantee permanence, production suitability, uptime, latency, model quality, or access from every region.
Resolve the live model route
This snapshot records 114 model IDs. Match the exact ID against the provider's live catalog before using it in code; zero-price routes and aliases can rotate while an older documentation page remains online. No stable base endpoint is published in this record, so use the linked provider documentation to identify the current request URL and protocol.
Confirm account and billing boundaries
An account is required, so public catalog evidence cannot by itself prove the limits, balance, or billing behavior attached to your workspace. The reviewed evidence explicitly says no payment card is required. Re-check the signup flow because eligibility and billing controls can change after the snapshot. The equivalent paid value shown here prices a documented allowance where possible; it is not cash, guaranteed savings, or protection from overage.
Re-read the prompt policy for your plan
This record classifies the reviewed handling as “Partially private”: No endpoint-wide public prompt-retention, training, operator-access, or deletion policy was found. Provider policies can distinguish free, paid, enterprise, opted-in, and feature-specific traffic, so verify the governing terms for the exact account and route that will receive your data.
Follow the evidence, then re-check it
The record links 5 first-party sources covering the offer, catalog, limits, pricing, terms, privacy, or adoption evidence available to the audit. Prefer the newest governing document or live catalog when sources disagree, and submit a correction when a provider changes a material term.
Match the quota to the workload shape
Translate the published allowance into the traffic pattern you actually expect instead of comparing headline totals alone. A daily token pool can look generous while a low requests-per-minute or concurrency ceiling blocks interactive bursts; a high request limit can still fail a long-context job when input, output, or per-request tokens are capped. Separate prompt tokens from generated tokens, include retries and tool calls, and test the largest realistic payload. If the provider documents more than one limit window, the tightest window at your peak load is the practical ceiling. Research, experimental, and community access can also carry eligibility or fair-use constraints that cannot be modeled as a simple number.
Plan fallback without changing the rules
A fallback should preserve more than API syntax. Confirm that the substitute route supports the required modality, context length, streaming behavior, structured output, tools, and safety controls, then compare its prompt-retention and training terms. An OpenAI-compatible request shape does not make providers operationally or contractually equivalent. Decide which errors may trigger a retry, cap retry storms, and prevent an exhausted free route from silently switching to a billable model. If deterministic output matters, record the model revision and sampling settings; rotating aliases and free-model pools can change behavior even when the endpoint remains available.
Monitor the offer as a dependency
Capture the model ID, response model field, rate-limit headers, usage fields, latency, HTTP status, and any provider request identifier for each test. Watch for authorization failures, quota exhaustion, catalog removal, policy revisions, and dashboard balance changes as separate failure modes. Re-check the provider’s live catalog and governing pages on a schedule proportionate to the workload’s importance, and keep the dated sources that supported your decision. Free capacity is especially suitable for prototypes, evaluations, fallbacks, and bounded workloads when the application can tolerate change; a production dependency still needs observability, an exit path, and an owner responsible for re-verification.
Test one complete request before scaling
Start with the smallest permitted request using the exact credential, model ID, endpoint, and account type you intend to deploy. Record the HTTP status, response headers, usage fields, latency, and any dashboard balance change. Then exercise the failure path: an invalid model, an exhausted quota, or a rate-limit response should fail clearly without silently switching to a paid route. If the provider supports streaming or tool calls, validate those features separately because a free model can expose a narrower capability set than its paid counterpart. Keep a budget ceiling outside the application whenever billing is possible, and avoid sending sensitive data until the observed route matches the reviewed data agreement.