Free-inference capacity expansion — 2026-10-21

What free inference the estate does not hold, which of it is real, and the exact human steps to claim it. One account per platform, registered by the owner under his own legal identity. No mass registration, no multi-accounting, no CAPTCHA automation, no aliases or throwaway mailboxes.

Claim labels used throughout:

VERIFIED — the vendor’s own page was read in this session and the sentence is quoted or summarised from it. The URL is given.
UNVERIFIED — not confirmed on a vendor page. Do not sign up on the strength of a summary.

Why this is a short list

The estate is already near-saturated on free seats, so the honest expansion is small. Two premises in the original brief failed verification outright: Cerebras has no permanently free tier, and Together AI has no free trial at all. Most of the “obvious” candidates (Groq, Google AI Studio, Hugging Face, Mistral, Cohere, NVIDIA NIM, Cloudflare Workers AI, ModelScope, Z.AI, Fireworks, DeepInfra, OpenRouter, Scaleway) are already held.

Held today

Read from the declared key names in acumen.env on tcp (names only; no value was read, printed, or copied) and cross-checked against the seat directory in the acumen catalog (config.go, extra.go). Held free-plan seats include:

Seat	Free mechanism	Notes
groq	free/low-cost fast chat	6M tokens/day community pool; 1000 RPM / 250k TPM on gpt-oss
google	AI Studio free tier	GEMINI_API_KEY + `_2` + `_3` all present
huggingface	router free credit	200k tokens/month
modelscope	free	6M tokens/day
mistral	free tier	1B tokens/month ceiling
cohere	trial	800k tokens/month
nvidia	NIM free credits	seat at integrate.api.nvidia.com, PlanFree
cloudflare	Workers AI free neurons	10k neurons/day, separate from tokens
zai	Z.AI / GLM free
fireworks	free credits
deepinfra	free credits
openrouter	`:free` model suffix	1M tokens/day
ollama-cloud	minimax free window	free until 2026-10-22 23:59 UTC, then paid
scaleway, nscale, inception, arcee, agentrouter, nararouter, xkiro, longcat, bazaarlink, bytez	signup credits	seat declared, plan prepaid

The limit figures in that table are the catalog’s own community free-tier table, not vendor pages re-read today. They are good enough to establish that a seat is held, which is all this section claims; they are not evidence for negotiating with a vendor.

Qwen is the exception that matters: the QWEN_* keys are present but dead, so Qwen capacity is effectively zero despite the env rows.

Candidate universe

Diffing the catalog (seat declared) against the env key names gives ten genuine candidates. Eight are declared seats with no key: sambanova, cerebras, together, siliconflow, hyperbolic, nebius, novita, ai21. Two have no seat at all: Alibaba Model Studio (DashScope) and OVHcloud AI Endpoints.

Ten researched, five survive as actionable. That is the finding, not a shortcoming of the research.

Priority order — best tokens-per-hassle first

#	Platform	Cost to get	Renews?	Card?	Human steps	Verdict
1	SambaNova SambaCloud	free	daily	no	email verify	best ratio — sign up
2	Alibaba Model Studio	free quota	no, 90d	no (inferred — see §2)	email + account info	biggest slug, one-time — sign up
3	OVHcloud AI Endpoints	US$200 trial credit (30d)	30d	yes	email + card + project	best throughput — sign up
4	Cerebras	$5 credit	no, 30d	yes	email + card	speed only — optional
5	OVHcloud anonymous	free	n/a	no	none	2 RPM, no account — wire as probe seat

1. SambaNova SambaCloud — free tier, renews daily, no card

VERIFIED — https://docs.sambanova.ai/docs/en/models/rate-limits.md

SambaNova defines the free tier by absence of a payment method: “Free Tier: Applied when there is no payment method linked with your account.” Developer Tier (with a card) is 20M tokens/day across all models; the free tier is far smaller but costs nothing and does not expire.

Free-tier limit	Value
RPM	20
RPD	20
TPD	200,000 (tokens per day, free tier only)

Model ids on the free tier, all production except the two previews: DeepSeek-V3.1, DeepSeek-V3.2 (preview), Meta-Llama-3.3-70B-Instruct, gpt-oss-120b, gemma-4-31B-it (preview). Twenty requests a day is the real constraint — this is a spillover/probe seat, not a bulk seat. 200k tokens/day is respectable if each request is long.

Signup: https://cloud.sambanova.ai — free account, API keys under “API keys and URLs”. Up to 25 keys per account. Keys cannot be viewed again after creation.

Fields:

Field	Value
Full name	James White
Email	owner’s own mailbox (fill at signup; must be reachable)
Company (optional)	Foxcore Cloud
Address (if asked)	7 Bell Yard, London WC2A 2JR
Payment method	do not add — adding a card moves the account to Developer tier and ends the free tier

Human step: email verification click. No card, no phone seen in the documented flow.

Wiring: already zero-code. The catalog declares this seat with KeyEnv SAMBANOVA_API_KEY, base https://api.sambanova.ai/v1, and a publishedSeat row carrying gpt-oss-120b and Meta-Llama-3.3-70B-Instruct at 128k context with price marked unknown. Only the key is missing.

2. Alibaba Model Studio — free quota per model, 90 days

VERIFIED — https://www.alibabacloud.com/help/en/model-studio/new-free-quota

First activation grants a free quota per eligible model, valid 90 days from activation, model release, or approval (whichever is later). Only models in the Singapore region whose deployment scope is International qualify. The quota is one-time per account: the page states that re-registering a new account does not grant additional quota, and unused quota is voided at expiry rather than paused.

The per-model token count is not in the public page text — the console shows it at activation. Record the number when it appears instead of assuming a figure.

Signup: https://modelstudio.console.alibabacloud.com (Singapore region) — requires completing account information before activation, which has been mandatory since 15 September.

Fields:

Field	Value
Full name	James White
Email	owner’s own mailbox
Company	Foxcore Cloud
Country/region	United Kingdom (International deployment scope)
Address	7 Bell Yard, London WC2A 2JR

Human step: email verification plus account-information completion (identity fields). Enable Free Quota Only (“worry-free mode”) in the console — it is off by default, and with it off the account starts billing pay-as-you-go the moment the quota is exhausted.

The “Card? no” entry in the priority table for this platform is an inference, not a vendor sentence: the free-quota page requires completing account information and says nothing about a payment method, and Alibaba Cloud International commonly asks for one during identity verification. Treat it as unverified — do not assume no card is needed until the console asks for one.

This is the direct fix for the dead QWEN_* keys: it buys working Qwen capacity for a defined window at no charge, and the expiry is known in advance.

Wiring: no seat exists. The catalog has an unrelated alibaba-token-plan entry in the disabled directory, which is not this product. A new seat is required — see “Seat registration” below. Confirm the OpenAI-compatible base host for the Singapore region in the console before writing it into the catalog; do not guess it.

3. OVHcloud AI Endpoints — US$200 one-month trial credit, 400 RPM per model, no usage cap

VERIFIED — https://docs.ovhcloud.com/en/guides/public-cloud/ai-machine-learning/ai-endpoints-capabilities.md and …/ai-endpoints-getting-started.md and …/ai-endpoints-billing.md

The rate limits are the best of any candidate by a wide margin:

Mode	Limit
Anonymous (no key)	2 requests/min per IP per model
Authenticated	400 requests/min per Public Cloud project per model

“There is no usage limit” beyond rate and payload size (2 MB body, 10 MB for vision-language models). This is a genuinely large throughput ceiling with no token accounting to fight.

Two things temper it. It is billed per token, not free — the free part is the US$200 first-project Public Cloud credit, which is real money that goes a long way at open-model prices. And an access key requires a card on the project: “Access keys created from Public Cloud projects in Discovery mode (without a payment method) cannot use the service.”

The US$200 figure is the World-market listing on https://www.ovhcloud.com/en/public-cloud/free-trial/ (“US$ 200 free credits for your cloud projects”); the en-gb page linked below shows £175. Either way it is a one-month trial credit, not a standing balance — the same page’s terms say the voucher “is valid for a period of one (1) month” and that its value “is listed in the publicly displayed currency for the market/country to which the Public Cloud account with the voucher is attached”. The expiry is tracked here the same way the Alibaba and Cerebras expiries are, and none of the three docs.ovhcloud.com pages cited above names the amount.

Signup: https://www.ovhcloud.com/en-gb/public-cloud/ai-endpoints/ then create a Public Cloud project; keys are generated in the control panel under AI Endpoints, and each key has an adjustable validity period.

Fields:

Field	Value
Account type	Individual or company (company: Foxcore Cloud)
Full name	James White
Email	owner’s own mailbox
Address	7 Bell Yard, London WC2A 2JR
Payment method	required for a key; the $200 credit is consumed first

Human step: email verification, card entry, and Public Cloud project creation. Highest friction of the list, highest ceiling.

Wiring: no seat exists. Base URL is documented as https://oai.endpoints.kepler.ai.cloud.ovh.net/v1 and the API is OpenAI-compatible. The catalog carries gpt-oss-120b in OVH’s catalogue; read the model list from the console rather than transcribing it from a blog.

Note the IAM caveat in OVH’s own docs: project administrators cannot see or revoke API keys created by added team members. One account, owner-held key, no team members — keep it that way.

4. Cerebras — $5 for 30 days, card required (optional)

VERIFIED — https://inference-docs.cerebras.ai/support/rate-limits and …/support/pricing

Cerebras is not a free tier in the renewable sense, and the brief’s “cerebras free” premise is wrong:

“Is there a permanently free tier? No.”
Free Trial = $5 in credits, expiring 30 days after grant, and it needs a verified payment method at signup; without one, “Playground and API access remain inactive”.
Free Trial limits: 5 RPM, 30k uncached TPM, 1M TPD on gpt-oss-120b and qwen-3.8-27b.

The only reason to bother is raw speed on a short evaluation. 1M TPD for 30 days is decent, and the card is not charged unless you buy credits — but the promo terms explicitly allow revocation for “fraud, abuse, or duplicate accounts”, so one account, no retries.

Signup: https://cloud.cerebras.ai

Fields: James White / owner’s mailbox / Foxcore Cloud / 7 Bell Yard, London WC2A 2JR. Card: yes.

Human step: email verify plus card verification (this may include a bank OTP — that is a human step, never automated).

Wiring: zero-code. The seat declares CEREBRAS_API_KEY at https://api.cerebras.ai/v1 and already carries a publishedSeat row sourced from https://inference-docs.cerebras.ai/models/overview (qwen-3.8-27b free-tier context 64k, with a note that paid is 128k/131k).

Cerebras models are also reachable through the already-held OpenRouter seat, which is the cheaper path if the only goal is a few Cerebras tokens. Do not create the account unless sustained Cerebras throughput is actually wanted.

5. OVHcloud anonymous — no account at all, 2 RPM

VERIFIED — same capabilities page as #3.

2 requests/min per IP per model, with no API key and no account. Useless for load, fine for liveness probes or a model-catalogue read from a host that needs one. Because acumen supports unauthenticated seats (AllowUnauthenticated), this can be wired without any credential at all. Treat it as an observability seat on the auto/probe path, never as chat capacity.

Rejected after verification

Do not spend time on these. Each was a brief assumption that a vendor page contradicts.

Candidate	Why rejected	Source
Together AI	“Together AI does not currently offer free trials. Access to the Together platform requires a minimum $15 credit purchase.”	https://docs.together.ai/docs/billing-credits.md
Cerebras (as a free tier)	No permanently free tier; $5/30d with card. Kept only as optional #4 above.	https://inference-docs.cerebras.ai/support/rate-limits
GitHub Models	Retired. The product page reads “GitHub Models has been retired.”	https://docs.github.com/en/github-models
SiliconFlow	Per-token pricing on the public model list; no free-tier or signup-credit statement found on the pricing page. Model prices match DeepSeek direct ($0.15/$0.60 per M), so there is no capacity argument over the seat already held.	https://www.siliconflow.com/models
Hyperbolic	Documentation and pricing describe GPU cloud (on-demand, reserved, private cloud), not a free inference tier. Seat declares `HYPERBOLIC_API_KEY`, which is absent from env.	https://docs.hyperbolic.xyz/

Not verified either way, so not recommended and not opened as tickets: Nebius Token Factory, Novita, AI21 Studio. Their catalog rows exist with free-quota hints in the community table, but no vendor page was read this session. Verify before signup.

Terms constraints (read before routing traffic)

Free tiers are not free capacity for customers. Three constraints apply to every seat above:

Purpose limits. Cerebras states the free trial is “for development, evaluation, and experimentation; not intended for production use.” Similar language governs most free tiers. Acumen should route free seats only to non-production classes; customer traffic belongs on prepaid or subscription seats.
One account per person. Every platform here allows one account per natural person or organisation. Duplicate accounts are grounds for revocation (Cerebras says so explicitly) and Alibaba states re-registration grants no extra quota. No second accounts, ever.
Expiry is not optional. Alibaba’s quota is voided at 90 days whether used or not; Cerebras’s at 30 days. Record the expiry date in the ticket when the account is created, and let the seat go dark rather than fall through to pay-as-you-go.

Seat registration in acumen

The brief’s proposed convention ACUMEN_<PLATFORM>_API_KEY is not what the catalog consumes. Two different systems are involved:

ACUMEN_* variables are acumen’s own knobs (ACUMEN_PROBE_EVERY, ACUMEN_EXACT_CACHE, ACUMEN_REQUIRE_KEYS, ACUMEN_ALLOWED_HOSTS). A vendor key stored under ACUMEN_* would be invisible to the catalog.
Each seat declares its own vendor-native KeyEnv, and that is the name the loader reads: SAMBANOVA_API_KEY, CEREBRAS_API_KEY, GROQ_API_KEY, HF_TOKEN (with HUGGINGFACE_API_KEY as a documented fallback), and so on.

So the wiring steps are:

Key placement. The operator appends NAME=value to acumen.env on tcp, root-owned, mode 0600. No agent writes this file, no value is printed, committed, or passed on a command line. Reference the path only.
Seat already declared (sambanova, cerebras, together): no code change. HasKey() moves the seat into the candidate set; restart acumen; confirm the seat and its models appear in the model list.
Seat absent (alibaba-model-studio, ovh-ai-endpoints): add an openaiSeat(id, base, keyenv, plan, features, notes) row in extra.go, add a freeQuota entry only where the tier is free rather than credit, and add a publishedSeat row sourced from the vendor page. That is a separate PR in tcpsystems/acumen with its own catalog test.
Verify the money. Send one cheap completion per new seat, then read the ledger: a free seat must settle at $0, a credit seat must decrement the credit. A seat that reports unknown price is not evidence of free — nominal() marks prices unknown by design, and budget then uses a conservative estimate.
Never widen a seat beyond its tier. Free seats get chat doors only, not embeddings, not responses, unless the vendor page says otherwise.

Tracker

Signup-ready platforms have an issue each, labelled capacity, listing the human steps in order. The two seats that need new catalog rows have one issue of their own. Nothing here is automated: every verification click is the owner’s.

Platform	Issue	Seat exists?
SambaNova SambaCloud	#38	yes, zero-code
Alibaba Model Studio	#39	no, needs seat
OVHcloud AI Endpoints	#40	no, needs seat
Cerebras (optional)	#41	yes, zero-code
OVHcloud anonymous	not ticketed — no signup involved	no, needs seat

The two seats that need new catalog rows are tracked together in #42.

Sources

All read 2026-10-21.