Skip to content
tcp.systems docs

Free-inference capacity — Western/US providers

Worker W101 (finder, Western/US) on model deepseek/deepseek-flash. Survey date 2026-10-21. Every claim below is labelled with how it was obtained:

  • (obs) observed in this session — a tool result on the URL named next to it.
  • (obs-3p) observed on a third-party page, not the vendor’s own docs. Treat as a lead, not a fact.
  • (unk) unknown. The signup app is client-rendered and the field list cannot be read without executing JavaScript, or no published number exists.

No account was created for this survey. No credential was minted, read or written. Signup-field lists are what the rendered page shows; where a form only exists after hydration, it says so.

Scope and skip list

Section titled “Scope and skip list”

The estate already holds these, via the acumen catalog and the environment key names — they were not surveyed:

deepseek, glm, qwen, minimax, gemini, grok, claude, codex, mistral, moonshot, opencode free.

mistral was on the requested list (La Plateforme free tier) and is skipped because we already hold it. nebius (Token Factory) is excluded as EU-based, outside this worker’s Western/US scope.

Surveyed: 17 platforms; 12 of them genuinely offer free inference today, 5 do not (kept in this document so nobody re-checks them next quarter).

Summary

Section titled “Summary”
PlatformFree inference todayHeadline quotaSignupCard for free tierVerdict
Groqyes, permanent200K TPD per model, 30 RPM, 1K RPDconsole.groq.comnotake
Cloudflare Workers AIyes, permanent10,000 neurons/day, 300 RPM text-gendash.cloudflare.comnotake
SambaNovayes, permanent200K TPD, 20 RPM, 20 RPDcloud.sambanova.ainotake (request-starved)
NVIDIA NIM (build.nvidia.com)yes, prototype tier~40 RPM, credit ceiling (unk)build.nvidia.comnotake, one-time
OpenRouteryes (:free variants)20 RPM, 50 RPD until $10 lifetime top-upopenrouter.ainotake
Modalyes — $30/month compute credit$30/mo, renews monthlymodal.com/signupnotake for batch
Cohereyes, trial key20 req/min/model, 1,000 calls/monthdashboard.cohere.comnotake, small
Novita AIyes — one $0 model + signup creditApodex 1.1 Mini free; $0.50 credit (obs-3p)novita.ainomarginal
Basetencredits only, no free tiercredit amount unpublished (obs-3p $30)login.baseten.colatermarginal
Together AItwo $0-priced endpointsTev1 4B Experimental (free output), Ternary Bonsai 27B ($0/$0)api.together.xyzunkmarginal
Hugging Faceno — PRO only$2/mo credits with PRO ($9/mo)huggingface.co/joinn/askip
Fireworks AI$1 starter credit only (obs-3p)no ongoing allowanceapp.fireworks.aiyes, for real useskip
Replicate“try for free” models, then billingno published creditreplicate.comyesskip
Cerebras$5 trial credit, 30-day expiry5 RPM, 30K uncached TPM, 1M TPDcloud.cerebras.aiyes, requiredskip
Hyperbolicno — inference gated behind $5 depositfree tier cannot call inference at allapp.hyperbolic.aiyes ($5)skip
DeepInfrano“add a card or pre-pay or you won’t be able to use our services”deepinfra.comyesskip
GitHub Modelsretired 2026-07-30nonen/an/adead

Dossiers

Section titled “Dossiers”

1. Groq

Section titled “1. Groq”
  • Free inference (obs). A Free plan with per-model limits, no expiry stated. From https://console.groq.com/docs/rate-limits: openai/gpt-oss-120b 30 RPM / 1,000 RPD / 8K TPM / 200K TPD; openai/gpt-oss-20b the same; qwen/qwen3.8-27b the same; meta-llama/llama-prompt-guard-2-* 500K TPD; Whisper 2K RPD with 28.8K audio-seconds/day. Limits are org-level. Cached tokens do not count against them. The page says the numbers shown are the Developer plan base and that the exact values live on the account’s limits page — so treat 200K TPD as the free-plan figure it is tabulated under.
  • Signup. https://console.groq.com/signup. Account type: individual developer; org is created on first login. Rendered signup is a single client-side B2B auth form (LoginB2B, form injected after hydration) — field list (unk). Third-party guides report email or Google/GitHub, email verification required, no phone, no card (obs-3p: getfreeai.net, freellm.net). No CAPTCHA was observed on the page shell; unk whether one appears inside the auth widget.
  • Why it ranks first. Nine-figure token allowance per model per day, no card, no phone, OpenAI-compatible base URL, and the strongest small models among the free tiers.
  • Cost of using it. 8K TPM is the binding constraint for long agent prompts; it needs client-side pacing, not just retries.

2. Cloudflare Workers AI

Section titled “2. Cloudflare Workers AI”
  • Free inference (obs). https://developers.cloudflare.com/workers-ai/platform/pricing/: “10,000 Neurons per day at no charge”, reset 00:00 UTC, on the Workers Free plan. Rate limits (obs, https://developers.cloudflare.com/workers-ai/platform/limits/): 300 requests/min for text generation, 3,000/min embeddings. Frontier models (kimi-k2.6/k2.7, glm-5.2/5.3, deepseek-v4-flash, deepseek-v4.1-flash, deepseek-v4-pro) require Workers Paid, so they are out of the free allocation.
  • What 10,000 neurons/day buys (derived from the published per-model neuron rates). @cf/meta/llama-3.2-1b-instruct 2,457 neurons/M input → ≈4.1M input tokens/day. @cf/openai/gpt-oss-20b 18,182 neurons/M input + 27,273/M output → ≈0.35M in + 0.23M out per day. @cf/qwen/qwen3-30b-a3b-fp8 is the best quality-per-neuron mid-tier at 4,625/M input.
  • Signup. https://dash.cloudflare.com/sign-up. Account type: Cloudflare account (email + password), then create a Worker and bind env.AI. Rendered form fields (unk) — Cloudflare’s signup is client-rendered. Cloudflare historically verifies by email link; Turnstile/CAPTCHA presence inside the signup flow is (unk). No card for the Workers Free plan (obs: the plans page distinguishes Free from Paid).
  • Why it ranks second. 300 RPM — ten times Groq’s request headroom — and no card. The tokens are cheap-model tokens, so it is bulk-classification capacity, not a substitute for a frontier model.
  • Estate fit. Already Cloudflare-adjacent (this repo’s console work), so the account is likely to exist for platform reasons before it is an AI account.

3. SambaNova Cloud

Section titled “3. SambaNova Cloud”
  • Free inference (obs). https://docs.sambanova.ai/docs/en/models/rate-limits: Free Tier applies when no payment method is linked. Per model: 20 RPM, 20 RPD, 200,000 TPD. Models: DeepSeek-V3.1, Meta-Llama-3.3-70B-Instruct, gpt-oss-120b, plus preview DeepSeek-V3.2, gemma-4-31B-it. OpenAI-compatible: https://api.sambanova.ai/v1 (obs, same page’s quickstart).
  • Signup. https://cloud.sambanova.ai → Login/Signup. Account type: individual. The landing page renders only the dashboard shell and a “Login/Signup” control; form fields (unk). No card for the free tier by definition (obs: the tier is defined as “no payment method linked”).
  • The catch (derived). 20 requests/day caps the request count even though the token allowance is 200K — that is 10K tokens per request, which only works for long, batched prompts. As a background agent lane it starves on requests; as a nightly batch summariser it is fine.
  • No expiry stated in the docs, and no card is requested to reach it.

4. NVIDIA NIM — build.nvidia.com

Section titled “4. NVIDIA NIM — build.nvidia.com”
  • Free inference (obs). https://build.nvidia.com/ front page states “Free inference with leading models” and lists Nemotron 3.5 Super VL, Nemotron 3.5 Lightning, DeepSeek-V4.1-Flash and Nemotron 3 Ultra as endpoint models. Each model page carries a “Prototype” block against https://integrate.api.nvidia.com/v1 with NVIDIA_API_KEY (obs, https://build.nvidia.com/deepseek-ai/deepseek-v4.1-flash.md) and is governed by the NVIDIA API Trial Terms of Service (obs, same page), which is the signal that this is a trial/prototype grant, not a production free tier.
  • Rate limit and credit ceiling. Search-indexed NVIDIA pages state “up to 40 requests per minute (RPM) for most models, with no per-token charge” (obs-3p). The commonly cited “1,000 free credits, 5,000 on request” is (obs-3p) only — neither number is on the pages this session fetched. Verify on first login before planning capacity.
  • Signup. https://build.nvidia.com → “Login”/“Get API Key”. Account type: NVIDIA account (NVIDIA Developer Program / NGC). Fields (unk); the historical flow is email + password + email verification, no card.
  • Estate fit. One-time credit ceiling means this is a spike resource, not a lane. Good for one-off evaluation of frontier open models (Nemotron 3 Ultra, DeepSeek V4.1 Flash) without paying.

5. OpenRouter (:free variants)

Section titled “5. OpenRouter (:free variants)”
  • Free inference (obs). https://openrouter.ai/docs/api_reference/limits: models whose id ends in :free get 20 requests/min and 50 requests/day while lifetime credits purchased are under 10; 1,000/day once $10 has been purchased at least once. Free-model limits do not apply to BYOK requests.
  • Signup. https://openrouter.ai — account via Google/GitHub/email; exact field list (unk). No card for the free tier.
  • Why it is on the list despite being an aggregator. It is the only route to several frontier-adjacent free models in one key, and the 1,000/day ceiling after a one-time $10 top-up is the cheapest way to multiply a free quota.
  • Caution. 50 requests/day is small; treat as an evaluation lane, and note that free-model availability churns without notice.

6. Modal (free monthly compute, not a token API)

Section titled “6. Modal (free monthly compute, not a token API)”
  • Free inference (obs). https://modal.com/pricing and https://modal.com/signup: Starter plan is “$0 + compute / month” with “$30 / month free credits”, 3 seats, 100 containers, 10 GPU concurrency. Credits renew monthly. Team is $250/mo with $100/mo credits.
  • What $30/month buys (derived from the published per-second rates on the same page). NVIDIA T4 at $0.000164/s ≈ 50 GPU-hours; L4 at $0.000222/s ≈ 37.5 GPU-hours; A100-40GB at $0.000583/s ≈ 14 GPU-hours. Enough to serve a small vLLM endpoint for a few hours a day, or to run batch evaluation jobs.
  • Signup. https://modal.com/signup. Account type: GitHub, Google or SSO — “Continue with GitHub”, “Continue with Google”, “Continue with SSO” are the only options (obs). No email/password form, no card for Starter.
  • Human verification. Whatever GitHub/Google enforces; Modal adds nothing observable on the signup page.
  • Estate fit. The only recurring free GPU compute on this list. It does not give a hosted model — it gives you somewhere to run one, which is what the estate wants for go test-adjacent batch work and evals.

7. Cohere (trial key)

Section titled “7. Cohere (trial key)”
  • Free inference (obs). https://docs.cohere.com/docs/rate-limits: evaluation (“trial”) keys are free but limited to 1,000 API calls a month; 20 requests/min per chat model (Command A+, A Reasoning, A Translate, A Vision, Command A, R+, R, R7B, North Mini Code); embeddings 2,000 inputs/min; rerank 10/min.
  • Signup. https://dashboard.cohere.com/welcome/register. Rendered fields (obs): Continue with Google, Continue with Github, Email, Password, Sign up — nothing else. Account type: individual.
  • Human verification. Email verification implied by the email/password path; a CAPTCHA is not rendered on the page (obs) and its absence inside the submit handler is (unk).
  • Estate fit. 1,000 calls/month is an evaluation budget. Useful because Command A / North Mini Code are model families we hold nowhere else, and rerank/embed are cheap to consume.

8. Novita AI

Section titled “8. Novita AI”
  • Free inference (obs, in part). https://novita.ai/pricing lists Apodex 1.1 Mini (256K context) at input free / output free. A $0.50 signup credit valid one year is (obs-3p) (pricepertoken, cheahjs list); the pricing page itself says only “Kostenlos zum Einstieg” (free to start).
  • Signup. https://novita.ai → sign up; account type individual; fields (unk). No card for the free model / signup credit per the third-party reports.
  • Estate fit. One free small model plus a tiny credit. Worth keying only because it is zero-cost and OpenAI-compatible; not worth engineering effort.

9. Baseten

Section titled “9. Baseten”
  • Free inference (obs). https://docs.baseten.co/organization/billing: “New workspaces receive credits for testing and deployment” and “Baseten doesn’t offer a separate free tier or perpetual free plan.” The credit amount is not published; (obs-3p) says $30. Credits are applied to the invoice, not redeemed as tokens.
  • Signup. https://login.baseten.co/. Rendered options (obs): Email (email link), Continue with Google, Continue with GitHub.
  • Verdict. One-time credits against a paid platform. Only interesting if we already need Baseten Model APIs; it is not capacity we can plan on.

10. Together AI

Section titled “10. Together AI”
  • Free inference (obs). https://www.together.ai/pricing shows no credit grant, and https://api.together.xyz/signup offers only “Continue with Google”, “Continue with GitHub”, “Continue with SSO” (obs). Two endpoints are priced at zero: Tev1 4B Experimental at $0.04 input / Free output, and Ternary Bonsai 27B at $0.00 / $0.00 (obs, same pricing page). Whether a zero-priced endpoint is callable without a payment method is (unk) — Together’s on-demand tiers generally expect a card.
  • Signup. Google / GitHub / SSO only; no email/password.
  • Verdict. Two free small-model endpoints, unclear gating. Lowest-effort way to find out is to log in with the existing GitHub account and call them.

11. Hugging Face Inference Providers

Section titled “11. Hugging Face Inference Providers”
  • Free inference (obs). https://huggingface.co/docs/inference-providers/pricing is explicit: Free users — monthly credits “None”. PRO gets $2.00/month; Team/Enterprise $2 per seat. Free users must purchase credits to use routed inference. A custom provider key avoids HF billing but then the provider’s own terms apply (i.e. it is that provider’s free tier, not HF’s).
  • Signup. https://huggingface.co/join. Account type: individual; fields (unk) (client-rendered). Email verification required in practice.
  • Verdict. Not a free source today. The hf-inference provider is largely CPU-class models, so even the PRO credit is small change. Skip unless a PRO seat is bought for other reasons.

12. Fireworks AI

Section titled “12. Fireworks AI”
  • Free inference (obs). https://fireworks.ai/pricing describes pre-paid, usage-based serverless billing and states no free tier. The $1 starter credit and “~10 RPM without a payment method” are (obs-3p) (morphllm, pricepertoken, cheahjs list); the pricing page itself shows only a free fine-tuning promotion (Muse Glimmer 30B SFT free until 11/30, no card).
  • Signup. https://app.fireworks.ai/signup. Rendered (obs): Email → Next, OR Sign up with Google / GitHub / LinkedIn. Account type individual.
  • Verdict. $1 is a test, not capacity. Skip unless we need Fireworks-specific serving paths.

13. Replicate

Section titled “13. Replicate”
  • Free inference (obs). https://replicate.com/docs/topics/billing: “You can run select models on Replicate for free, but after a bit you’ll be asked to set up billing.” No published credit amount, no daily allowance. Selected free models live in https://replicate.com/collections/try-for-free.
  • Signup (obs). https://replicate.com/signin offers exactly one path: Sign in with GitHub. So the account type is “GitHub account”; the field list is whatever GitHub returns; Replicate adds no form of its own.
  • Verdict. Free only until the trial nudge, then a card. Fine for a one-off image/audio model evaluation, useless as a lane.

14. Cerebras

Section titled “14. Cerebras”
  • Free inference (obs). https://inference-docs.cerebras.ai/support/rate-limits Free Trial: gpt-oss-120b and qwen-3.8-27b at 5 RPM, 30K uncached TPM, 1M TPH, 1M TPD — but “New accounts receive $5 in free credits after adding a verified payment method”, credits expire 30 days after grant, and “If you skip adding a payment method at sign-up, Playground and API access remain inactive until you do.” The FAQ also states there is no permanently free tier.
  • Signup. https://cloud.cerebras.ai — client-rendered SPA, fields (unk).
  • Verdict. Excellent tokens-per-second hardware, but the entry price is a verified card plus 30-day expiry. It fails the “free inference today without a card” test; keep it on the paid-evaluation list instead.

15. Hyperbolic

Section titled “15. Hyperbolic”
  • Free inference (obs). https://www.hyperbolic.ai/docs/general/account-management: “You may not launch GPU instances nor storage volumes, or use Inference, with a free-tier account.” $5 one-time deposit promotes to Pro. Also shows a phone-verification credit banner in its own screenshot.
  • Verdict. Not a free-inference provider. Closed for this purpose.

16. DeepInfra

Section titled “16. DeepInfra”
  • Free inference (obs). https://deepinfra.com/pricing: “You have to add a card or pre-pay or you won’t be able to use our services.” No free tier stated.
  • Verdict. Card-gated. Cheap per token, but not free capacity.

17. GitHub Models

Section titled “17. GitHub Models”
  • Retired (obs). https://docs.github.com/en/github-models/use-github-models/prototyping-with-ai-models: “As of July 30, 2026, GitHub Models has been fully retired. The playground, model catalog, inference API, and bring your own key (BYOK) are no longer available to any customer.” Redirects to Azure AI Foundry / Copilot.
  • Verdict. Dead. Any internal note still listing GitHub Models as a free source is stale and should be corrected.

Ranking by tokens per unit of hassle (derived)

Section titled “Ranking by tokens per unit of hassle (derived)”

Hassle = a card, a phone number, a CAPTCHA, an expiring credit, or a per-day request cap that starves a background agent.

  1. Groq — 200K TPD per model, 30 RPM, no card/phone, OpenAI-compatible. Best raw free allowance on the list.
  2. Cloudflare Workers AI — 10K neurons/day ≈ 0.35M–4.1M tokens/day depending on model, 300 RPM, no card, and the account likely already exists.
  3. SambaNova — 200K TPD with no card, but 20 requests/day forces batching; rank it third on volume and first on “long single prompt per call”.
  4. Modal — $30/month recurring credits (≈14–50 GPU-hours) for self-hosted inference; highest ceiling, highest setup cost.
  5. OpenRouter — 50 RPD free, 1,000 RPD after a one-time $10.
  6. Cohere — 1,000 calls/month, strong non-held model families.
  7. NVIDIA NIM — free prototypes with a credit ceiling; spike resource.
  8. Novita — one free small model, no effort to hold.
  9. Baseten / Together / Fireworks — one-time or unclear credits.
  10. Replicate / HF / Cerebras / Hyperbolic / DeepInfra — card-gated or empty.

Recommendation: open accounts only for 1–3 (and 4 if we want recurring GPU hours). Those three together give roughly 400K free hosted tokens/day plus a 300 RPM bulk lane at zero card risk — more than any single paid seat outside DeepSeek Flash, and enough to move a slice of the free-tier routing off the models we already pay for.

Verification still owed

Section titled “Verification still owed”
ItemWhy it is unknownHow to close
Groq, Cloudflare, Together, Baseten, SambaNova, Cerebras signup field listsclient-rendered auth widgetsread the form once in a browser, or create the account
CAPTCHA / phone at Groq and Cloudflare signupinjected after hydrationsame
NVIDIA credit ceiling (1,000 / 5,000) and 40 RPMonly in NVIDIA pages this session did not fetchlog in, read the account’s usage page
Together zero-priced endpoints callable without cardpricing page does not saycall Tev1 4B Experimental with a fresh key
Baseten signup credit amountdocs say “credits” without a numberbilling dashboard after signup
Novita $0.50 credit and 1-year validitythird-party onlybilling dashboard
Fireworks $1 starter creditthird-party onlybilling dashboard

Evidence index

Section titled “Evidence index”

Primary pages read in this session (all webfetch, 2026-10-21):

Third-party trackers used only for leads (never as the sole basis of a claim marked obs): pricepertoken.com, benchlm.ai, morphllm.com, the cheahjs free-llm-api-resources list, getfreeai.net, freellm.net.

Issue index

Section titled “Issue index”

One issue per platform, label capacity, in tcpsystems/ops, opened 2026-10-21.

PlatformIssuePlatformIssue
Groq#44Baseten#52
Cloudflare Workers AI#45Together AI#53
SambaNova#46Hugging Face#54
NVIDIA NIM#47Fireworks AI#55
OpenRouter#48Replicate#56
Modal#49Cerebras#57
Cohere#50Hyperbolic#58
Novita AI#51DeepInfra#59
GitHub Models (dead)#60