**Final shortlist is OVERSEAS ONLY.** Chinese domestic platforms
(智谱 BigModel/Z.ai, 硅基流动 SiliconFlow, 火山引擎豆包 Ark, 阿里云百炼 Bailian,
百度千帆 Qianfan, DeepSeek, Kimi/Moonshot, 腾讯混元 Hunyuan, plus 讯飞 Spark)
are **excluded from the recommendation**, even though the first research pass
found several of them very strong on Chinese. They are listed only in
“Excluded — do not use for this brief” so the parent agent can see what was
deliberately left out.
**“Overseas model” is applied strictly here:** the model maker must be
non-Chinese. That matters, because several overseas *hosts* serve excellent
Chinese-origin models — above all Alibaba Qwen on Groq, DeepSeek on
SambaNova/Chutes, and Z.ai GLM on OpenRouter/HF. Those models are flagged
`[Chinese-origin — excluded under the strict rule]` wherever they appear.
If the parent relaxes the rule to “overseas *host* is enough”, Qwen 3.8/3.6
on Groq would immediately become the best Chinese option after Gemini;
under the strict rule it is not in the shortlist.
**Use case arithmetic.** “A few dozen requests/day, normal use” is taken as
~20–60 requests/day, spread out, not bursty. “Several thousand Chinese
characters” of system prompt is roughly 2,000–5,000 tokens (GLM docs give
~1 token ≈ 1.6 Chinese characters; Kimi docs give ~1.5–2 characters/token),
so it fits every context window below by a wide margin. The binding
constraints are therefore **requests/day, tokens/minute and expiry**, not
context size — except where noted.
**Confidence / dating.** All figures below are from web-index sources read
4 Nov 2026 (UTC), cross-checked against Oct–Nov 2026 trackers. Free tiers
churn monthly (GitHub Models retired July 2026 is the warning example).
Treat every number as “check the console before shipping”, especially
Google, which no longer publishes static free-tier RPM/RPD.
For a server-side backend, overseas-only, Chinese-heavy, long system prompt,
low volume:
1. **Google AI Studio / Gemini — first choice on quality, if the server's
egress region is supported.** Flash-Lite/Flash, 1M context, OpenAI-compatible
endpoint, no card for the free tier in most regions, and by far the best
overseas Chinese. Its hard problems are regional availability
(mainland China, Hong Kong and Macau are NOT on the Gemini API list) and
unpublished, shrinking daily caps — use Flash-Lite, not Flash, for volume.
2. **OpenRouter `:free` (Nemotron / Gemma 4) — best fallback / second key.**
No card to start, 20 RPM / 50 RPD (1,000 RPD after a one-time ≥$10 credit
purchase), 128K–1M context, fully OpenAI-compatible. Chinese is
“supported on paper” for both Gemma 4 and Nemotron, not native-grade.
3. **Requesty free models — best daily headroom among aggregators.**
200 requests/day, no card, OpenAI-compatible, same overseas families
(Gemma 4, Nemotron). Use it as the fallback behind Gemini/OpenRouter.
4. **Groq `gpt-oss-120b` — fastest, most generous daily cap, weakest Chinese
of the top four.** 30 RPM / 1,000 RPD / 8,000 TPM / 200,000 TPD, 131K
context. The 8K TPM cap is the real long-prompt trap: a ~4K-token
system prompt allows roughly one such call per minute.
5. **Cloudflare Workers AI — permanent, privacy-strong, but Llama's Chinese
is officially unsupported.** 10,000 Neurons/day, no card, 128K context.
Do **not** build the primary backend on: GitHub Models (dead), Together
(no free tier any more), Cerebras (card-required 30-day trial only),
Hugging Face alone ($0–$0.10/month — effectively not a free source),
Pollinations classic / LLM7 / small relays (emergency fallback only),
or any single free tier without a second provider key.
---
· **How to get a key:** Google account → AI Studio → API key. No credit
card for the free tier in most regions. Sources conflict on
EEA/UK/Switzerland: one tracker says billing must be enabled even for
free-eligible models there; check the console for the account's region.
· **Card:** No (outside that EEA/UK/CH caveat).
· **What is free:** Permanently rate-limited free tier, not one-time credit.
Free input + output tokens on Flash / Flash-Lite models. Paid tier does
not train on prompts; **free tier prompts may be used to improve Google
products** — do not send sensitive/unpublished system-prompt IP without
accepting that.
· **Quota / rate limits:** Google stopped publishing static free-tier
numbers in Dec 2025; limits are shown per-project in AI Studio.
Community reports, Sep 2026: Gemini 3.x Flash ~20 RPD, Flash-Lite
~500 RPD. Older published figures (2.5 Flash 10–15 RPM / 250–1,500 RPD)
are stale after the Dec 2025 cuts (reported 50–80% cut). **Conflict to
flag:** Google's pricing page still listed 2.5 Pro as “Free of charge”
in Oct 2026, while multiple trackers say Pro moved to paid-only in
Apr 2026. Do not plan on Pro being free.
· **Context:** 1M tokens for current Flash / Flash-Lite (1,048,576).
Comfortably fits a several-thousand-character system prompt.
· **OpenAI-compatible:** Yes, via
`https://generativelanguage.googleapis.com/v1beta/openai/` with the
standard OpenAI SDK (`/chat/completions`). Native API is separate.
· **Free models:** Gemini 3.8 / 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite,
3 Flash Preview, 2.5 Flash / Flash-Lite, Gemma 4. NOT free: Gemini 4
Argon, Gemini 3.1 Pro Preview, image/video models.
Flagship-grade? Flash is Google's current workhorse tier, not the top
Pro tier — but it is a frontier proprietary model, a class above the
small open-weight free models elsewhere.
· **Chinese support:** **Best overseas option.** Gemini is a multilingual
flagship; independent work shows Gemini Flash Chinese accuracy close to
its English accuracy on judgement tasks (83.3% ZH vs 87.5% EN in one
LLM-as-judge study) and strong multilingual maths. For short Chinese
Q&A and ZH↔EN translation this is the safest overseas choice. Test
translation style on the site's own sentences — no overseas open-weight
model matches it for idiomatic Chinese.
· **Gotchas / region — decisive:** The Gemini API / AI Studio region list
(~230 territories) **excludes mainland China, Hong Kong and Macau**
(Taiwan is listed). A call from an unsupported egress IP returns
`400 FAILED_PRECONDITION: User location is not supported`. If the
backend server is in mainland China / Hong Kong, Gemini is not a
compliant direct option — the server's exit region, not the developer's
location, decides. Also: no SLA on free tier, roster/limits change
without notice, 2.5 models restricted to prior users since Sep 2026.
· **Sources:**
- https://www.morphllm.com/gemini-api-pricing (pricing page mirror, Oct 2026)
- https://agentdeals.dev/gemini-api-pricing-2026 (Dec 2025 cuts; 20 RPD Flash / 500 RPD Flash-Lite, Sep 2026)
- https://www.21cloudbox.com/support/google-gemini-china.html (region list, Oct 2026)
- https://ipjudge.org/en/guides/gemini-api-location-not-supported/ (API list excludes mainland China/HK/Macau)
- https://learn.microsoft.com/en-us/azure/api-management/openai-compatible-google-gemini-api (OpenAI-compat base URL)
· **Status:** Fully retired **30 July 2026**. Playground, catalog,
inference API and BYOK are gone for all customers, including existing
ones; closed to new customers 16 June 2026, brownouts 16/23 July 2026.
Any guide still recommending a GitHub PAT as a free LLM key is out of date.
· Historical free limits (for migration context only): low tier 15 RPM /
150 RPD / 8,000 input / 4,000 output tokens per request; high tier
10 RPM / 50 RPD. That 8K/request cap would in any case have been tight
for a long system prompt. Successor is Microsoft Foundry Models (paid,
Azure subscription) — not a free tier.
· **Source:** https://learn.microsoft.com/lv-lv/azure/foundry/foundry-models/how-to/quickstart-github-models
· **How to get a key:** Email signup at console.groq.com → API Keys.
No card on the Free plan.
· **Card:** No.
· **What is free:** Permanent rate-limited Free plan, not expiring credit.
Developer tier (card, $0 minimum spend) raises limits ~10× and gives a
25% token discount; that is pay-per-token, not free.
· **Quota / rate limits (per chat model, Sep–Oct 2026):**
30 RPM, 1,000 RPD, 8,000 TPM, 200,000 TPD for `gpt-oss-120b`,
`gpt-oss-20b` and `qwen/qwen3.8-27b`. Llama 3.1 8B / 3.3 70B left the
free and Developer tiers 16 Aug 2026 (enterprise-only). One Oct 2026
record notes `gpt-oss-safeguard-20b` was cut to 3 RPM / 2,000 TPM
without announcement — verify the exact model row in the console.
· **Context:** 131,072 tokens (gpt-oss-120b/20b, Qwen 3.8). Fits the long
prompt easily on paper.
· **OpenAI-compatible:** Yes, `https://api.groq.com/openai/v1`.
· **Free models:** OpenAI `gpt-oss-120b` / `gpt-oss-20b` (open-weight,
120B MoE, strong code/reasoning) plus `[Chinese-origin — excluded]`
`qwen/qwen3.8-27b`. No flagship proprietary model is free here.
· **Chinese support — the honest caveat:** Under the strict overseas-model
rule, Groq's Chinese story is its weakest point. OpenAI's own model card
gives gpt-oss-120b MMMLU Chinese 83.6 at high reasoning, but an
independent evaluation found both gpt-oss variants **below 30% on
C-Eval** (native Chinese exam benchmark) and “severe weakness in
multilingual evaluation”, vs Qwen 3 far higher. Reconciliation: it can
handle short Chinese chat and straightforward ZH↔EN translation, but
expect literal/awkward phrasing and weaker Chinese knowledge than
Gemini, and test it on the site's actual prompts before committing.
If Qwen were allowed, Groq+Qwen would be a top-2 Chinese option.
· **Long-prompt trap:** 8,000 TPM. A single request carrying a
~3,000–5,000-token system prompt is fine alone; two such requests in
the same minute, or one request plus a long history, can 429. Space
calls out, and/or use prompt caching where offered (cached tokens
reportedly do not count against the cap on some models — verify).
· **Privacy:** Groq's agreement does not permit training on inputs/outputs
without permission; it may log for up to 30 days for abuse/troubleshooting,
with a Zero Data Retention toggle.
· **Sources:**
- https://toolfreebie.com/groq-vs-cerebras-vs-gemini/ (Sep 29 2026 test)
- https://klymentiev.com/blog/groq-pricing (Sep 2026, Llama removal)
- https://agentdeals.dev/vendor/groq (Oct 2026, safeguard cut, privacy)
- https://arxiv.org/html/2508.12461v2 (gpt-oss C-Eval <30%)
- https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf (MMMLU Chinese 83.6)
· **How to get a key:** Account on Mistral Studio/La Plateforme, create
an API key. No card in Free mode.
· **Card:** No for Free mode.
· **What is free:** “Free mode” — recurring included monthly usage, amount
**not published**. The pricing page's former “$10/mo in API credits”
line disappeared 7 Oct 2026 with no announcement; the only reliable
number is the account's own Limits page. This is a prototyping tier,
not a published quota.
· **Quota / rate limits:** Not published numerically. Free mode is
described as the lowest limits, for evaluation/prototyping.
· **Context:** 128K for the free-eligible small models (Nemo, Ministral);
Mistral Small/Medium/Large are 128K–256K but see below.
· **OpenAI-compatible:** Yes.
· **Free models — critical narrowing:** Tested 29 Sep 2026 on a free key,
`open-mistral-nemo`, `open-mistral-7b`, `codestral-latest` and
`ministral-3b/8b/14b-latest` answered; `mistral-small-latest` and all
Medium models returned 429 with a 0/min limit, and Large returned 403.
In other words **the free tier does not include Mistral's capable
Small/Medium models** — only small open models. Do not shortlist
Mistral on the strength of Medium/Large benchmarks.
· **Chinese support:** Mistral's larger models list Chinese among
supported languages (Small 24B: Chinese MMLU 70.35 in its model card),
and Medium 3.5 lists Chinese/Japanese/Korean. But the *free* models are
the small ones, so expect serviceable short Chinese Q&A and basic
translation, below Gemini and below a native Chinese model, and
weaker on long, instruction-heavy system prompts. Test, do not assume.
· **Gotchas:** Consumer Vibe/Le Chat subscription and API billing are
separate; a Pro subscription buys no API credits (older sources that
say otherwise are stale). Free roster and the $10 credit line have
already changed twice in 2026.
· **Sources:**
- https://www.aifreeapi.com/en/posts/mistral-api-free-tier (Oct 9 2026, $10 line removal)
- https://toolfreebie.com/mistral-free-api/ (Sep 29 2026 free-key test)
- https://docs.api.nvidia.com/nim/reference/mistralai-mistral-medium-3-5-128b (Chinese listed, 256K — paid model, for language evidence only)
· **Status (Aug–Oct 2026, multiple trackers):** Free tier / $25 signup
credit **discontinued**. Platform is prepaid with a **$5 minimum
credit purchase** to access it. Older guides still quoting $25 are stale.
· **What remains:** Some models are listed at $0/M tokens on the pricing
page (e.g. a small open model in one Sep 2026 report), but Together's
own support docs are quoted as saying there is no free trial and no
free tier, and access requires the $5 purchase. Treat it as
**not a free option** for this brief; as a paid fallback it is
OpenAI-compatible (`https://api.together.xyz/v1`), 200+ open models,
and cheap per-token rates for Llama/Qwen-class models — but those
cheap routes include Chinese-origin models, so the overseas-only
shortlist would use its Llama/Gemma-class routes only.
· **Sources:**
- https://www.spheron.network/blog/free-gpu-cloud-credits-2026/ (Aug 2026: discontinued)
- https://aicost.tools/blog/free-ai-compute-tiers-trials-grants/ (Sep 2026: $5 minimum, no free credit)
· **How to get a key:** Free HF account (email) → token. No card to
create the account/token.
· **Card:** No to sign up; card needed to *buy* credits beyond the
(tiny/none) allowance.
· **What is free — conflicting sources, report both:** Sep 2026 trackers:
$0.10/month Inference Providers credit on free accounts, $2/month on
PRO ($9/month). 21 Oct 2026 check: **free accounts get no monthly
credits**; PRO gets $2; free users must buy credits or upgrade.
Either way this is **not a usable free tier for a daily backend**:
$0.10 buys a handful of frontier calls a month. The real free value
of HF is as a **BYOK router** (see Aggregators): plug in your own
Groq/OpenRouter/etc. key and HF does not charge for the call.
· **Quota:** Provider-side RPS/concurrency, not a published HF request cap;
the credit is the cap.
· **Context:** Depends on the routed model/provider (typically 128K for
Llama-class, 131K for gpt-oss).
· **OpenAI-compatible:** Yes for chat: `https://router.huggingface.co/v1`
(`/chat/completions`). Other tasks need the HF SDKs.
· **Free models:** None priced at $0 as a standing offer; routed models
are billed at partner rates. HF's own `hf-inference` is mostly CPU,
no chat models since July 2025.
· **Chinese support:** Inherits the routed model's Chinese — i.e. Gemma/
Llama-class overseas models are moderate (see §5), Qwen/DeepSeek routes
are Chinese-origin and excluded under the strict rule.
· **Sources:**
- https://aifree.dev/learn/free-hugging-face-inference (Oct 21 2026: no free credits)
- https://klymentiev.com/blog/huggingface-inference-api (Sep 11 2026: $0.10/month)
· **How to get a key:** Account, no payment method linked = Free tier.
(Note: SambaNova's Plans page wording conflicts with its rate-limit
docs on whether a payment method is needed; the rate-limit docs'
“no payment method linked” definition is the one trackers cite.)
· **Card:** No (on the Free-tier definition).
· **What is free:** Permanent daily allowance, not expiring credit:
**20 RPM, 20 requests/day, 200,000 tokens/day per model**.
· **Context:** 128K for Llama 3.3 70B, gpt-oss-120b, DeepSeek-V3.1;
DeepSeek-V3.2 preview 32K; Gemma 4 31B preview 128K.
· **OpenAI-compatible:** Yes, `https://api.sambanova.ai/v1`.
· **Free models:** `Meta-Llama-3.3-70B-Instruct`, `gpt-oss-120b`,
plus preview `DeepSeek-V3.2` `[Chinese-origin — excluded]` and
`gemma-4-31b-it`. MiniMax models are Developer-tier only.
· **Chinese support — weak, and this is a fit warning:** Meta's Llama 3.3
model card lists supported languages as English, German, French,
Italian, Portuguese, Hindi, Spanish and Thai — **Chinese is NOT a
supported language**, and Meta discourages use in non-supported
languages without fine-tuning. Stock Llama 3.3 will answer in Chinese
(it has seen Chinese text) but with the known “Chinese question,
English/mixed answer” failure mode of base Llama instruct models.
Gemma 4 (140+ languages incl. Chinese) is the better SambaNova Chinese
route, but it is a preview model that “may be removed at short notice”.
· **Fit for this brief:** 20/day per model is **below “a few dozen”
if the site uses one model**; splitting across Llama + gpt-oss + Gemma
could arithmetically reach 60/day, but that is fragile and the models
differ in Chinese quality. Treat as a tertiary fallback, not a primary.
· **Source:** https://blog.laozhang.ai/en/posts/free-ai-api-tiers-compared (Sep 26 2026, official rate-limit docs)
· **Language source:** https://docs.api.nvidia.com/nim/reference/meta-llama-3_3-70b-instruct (supported languages list)
· **How to get a key:** Account + **verified payment method required**,
even for the trial.
· **Card:** **Yes, required.**
· **What is free:** $5 credit, **expires 30 days after grant** (from
16 July 2026). “There is no permanently free tier” (Cerebras docs,
Oct 2026). After that, pay-as-you-go.
· **Quota:** Trial: 5 RPM, 30K uncached TPM / 90K total TPM, 1M tokens/day
on `gpt-oss-120b` and `qwen-3.8-27b` `[Chinese-origin — excluded]`.
· **Context:** **65K on the free trial** (131K paid) for gpt-oss-120b.
65K still fits a several-thousand-character prompt, so context alone
does not disqualify it — expiry + card does.
· **OpenAI-compatible:** Yes, `https://api.cerebras.ai/v1`.
· **Chinese support:** Same gpt-oss caveat as Groq (§3). Very fast
(~3,000 tok/s) but that does not help Chinese quality.
· **Source:** https://agentdeals.dev/vendor/cerebras (Oct 26 2026, quoting Cerebras rate-limit docs)
· **How to get a key:** Cloudflare account, no card. Workers AI is called
from a Worker or via REST; no separate “API key” in the classic sense
for the free allowance.
· **Card:** No.
· **What is free:** **10,000 Neurons/day, reset 00:00 UTC**, on free and
paid Workers plans alike. A Neuron is a compute unit, not a token:
e.g. Llama 3.1 8B fp8-fast ≈ 4,119 Neurons/M input, 34,868/M output
(≈287K output tokens/day on that model alone); Llama 3.3 70B fp8-fast
≈ 26,668/M input, 204,805/M output (≈49K output tokens/day).
On the free Workers plan you stop with error 3036 when it runs out;
overage billing ($0.011/1,000 Neurons) needs the $5/month Workers Paid plan.
· **Quota:** Text generation ~300 RPM account-wide (20 RPM on larger
frontier models, per one tracker); the Neuron budget binds first.
Worked example from a tracker: 30 chats/day at 2,000 input + 400 output
tokens on Llama 3.3 70B ≈ 4,060 Neurons/day — inside the free budget.
A long system prompt is input-heavy, which is the *cheap* side on
Neurons, so this use case fits Workers AI better than output-heavy ones.
· **Context:** 128K for Llama 3.3 70B / 3.1 8B.
· **OpenAI-compatible:** Workers AI has an OpenAI-compatible surface;
the primary interface is the Workers binding / REST API — verify the
exact endpoint for the chosen model before wiring a generic client.
· **Free models (overseas):** Llama 3.3 70B, Llama 3.1 8B, Llama 4 Scout,
Gemma 3/4, Mistral Small 3.1, gpt-oss-120b/20b, Nemotron 3.
`[Chinese-origin models also hosted — Qwen, Kimi, GLM, DeepSeek —
excluded under the strict rule.]`
· **Chinese support:** Same Llama limitation as SambaNova (§7) for the
70B route; Gemma 4 is the better Chinese route on this host (140+
languages). Cloudflare's privacy stance is the strongest here:
explicit commitment not to train on customer prompts/outputs without
consent.
· **Sources:**
- https://www.logeshwaran.org/2026/10/cloudflare-workers-ai-pricing-free.html (Oct 2026, Neuron math)
- https://klymentiev.com/blog/free-llm-api (Sep 2026, 10K Neurons/day, model list)
- https://dev.to/akaranjkar08/personal-ai-agent-on-cloudflare-free-tier-2026-code-2i8i (Oct 2026, worked example)
---
Aggregators are not model makers. They add one key, routing, fallbacks
and caching, and they subtract: an extra hop that sees the prompt, a
second set of terms, and a free roster that is subsidised and can vanish.
**Rule for this backend: an aggregator is a fallback, never the only key.
Never send the long system prompt (if it contains unpublished product/
business logic) through a relay whose training/retention terms you have
not read for that exact model route.**
#### A1. OpenRouter — USA (also the §1-requested free-model source)
· **Register / card:** Account + API key; **no card to start**.
Buying credits needs payment; the 5.5% fee (min $0.80) is on top-up,
not per token.
· **Free quota:** `:free` models: **20 RPM, 50 RPD**; **1,000 RPD once
lifetime credit purchases reach $10** (a lifetime threshold, not a
balance — spend it down and the 1,000/day stays; a negative balance
blocks even free models with 402). No token/day cap at platform level;
upstream provider saturation 429s happen anyway.
· **OpenAI-compatible:** Yes, `https://openrouter.ai/api/v1`. Also
`openrouter/free` router picks a free model at random (do NOT use it
for a deterministic backend — model changes per call).
· **Free roster (churn — verify live):** 15–19 text `:free` models in
Oct–Nov 2026, led by **NVIDIA Nemotron 3 / 3.5 (128K–1M context)** and
**Google Gemma 4 26B/31B (262K)**, plus Cohere North Mini Code
(256K) and others. DeepSeek R1, Qwen3 Coder 480B, Llama 3.3 70B and
Llama 4 Scout are **no longer free** here. A 3 Nov 2026 daily probe
found only 7 of 14 listed free models answering that day; Gemma 4
routes were 429ing from Google AI Studio for 7 days straight.
· **Chinese:** Gemma 4: 140+ languages incl. Chinese (Google open-weight,
decent but not native-grade; one tracker notes Gemma 4 needs Mandarin
fine-tuning for best results on Chinese legal/financial/medical text).
Nemotron 3: NVIDIA lists Chinese as a supported language
(EN/FR/DE/IT/JA/ES/ZH, 1M context). Both are usable for short chat and
straightforward ZH↔EN; expect Gemini to beat both on idiom and
long-instruction following. **Privacy trap:** OpenRouter itself does
not store prompts unless you opt in, but the *serving provider* may
train/retain — 9 of 19 free models in an Oct 2026 audit ran on hosts
marked “may train”. Turning on “deny training” for free models can
leave *no* endpoint for a given model. Set
`provider: {data_collection: "deny"}` / ZDR only if you accept fewer
routes, and test the exact model.
· **Stability / runaway risk:** **Low–medium.** Established company,
paid business alongside free tier, public status and per-key usage
API (`GET /v1/key` shows free-model daily use). Risk is roster churn
and upstream 429s, not disappearance overnight — but there is no SLA
on free models.
· **Sources:** https://klymentiev.com/blog/openrouter-free-tier (Nov 3 2026 probe);
https://www.teamday.ai/blog/best-free-ai-models-openrouter-2026 (Oct 28 2026 catalog);
https://www.truefoundry.com/blog/openrouter-free-models-whats-actually-free-and-what-it-costs-you (Oct 5 2026 data-policy audit)
#### A2. Requesty — UK
· **Register / card:** Free account + API key; **no card, no trial expiry**
on the Free plan.
· **Free quota:** **200 requests/day shared across free models** —
4× OpenRouter's unfunded cap, and the single best *daily* headroom
among no-card aggregators. No published RPM in the sources found;
verify in the dashboard.
· **OpenAI-compatible:** Yes, `https://router.requesty.ai/v1`, plus
Anthropic-format endpoint. Routing, caching, fallbacks, spend tracking,
EU data residency included on free plan. Paid: provider rate + 5%.
· **Free models:** 12–13 at $0 (Sep–Oct 2026): Nemotron 3 Ultra 550B,
Nemotron 3 Super 120B, Nemotron 3.5 Lightning, Gemma 4 31B, Leanstral,
Muse Glimmer, Ling-class models, etc. (Some listed free models are
Chinese-origin — DeepSeek routes — excluded under the strict rule;
the overseas routes above are the usable ones.)
· **Chinese:** Same Gemma/Nemotron verdict as OpenRouter. **Privacy flag
from a Sep 2026 audit:** Requesty's list marks the free NVIDIA-hosted
models (Nemotron, Muse Glimmer) as retaining inputs 30 days and using
them for training under NVIDIA API Trial terms; Gemma 4 31B and
Leanstral are marked not used for training. If the system prompt is
sensitive, prefer the non-training routes and verify the current
per-model terms.
· **Stability / runaway risk:** **Low–medium.** Real company, paid
catalogue (600+ models) on the same key, so the free tier is a funnel
rather than the whole business — materially safer than a pure free
relay. Still no SLA on free models; roster can change.
· **Sources:** https://www.requesty.ai/free-models ; https://docs.requesty.ai/quickstart ;
https://github.com/mvalentsev/awesome-free-ai-coding/blob/HEAD/providers/requesty.md (Oct 26 2026 probe)
#### A3. Vercel AI Gateway — USA
· **Register / card:** Vercel account. **Conflicting sources on card:**
Sep 2026 changelog says $5/30 days, no card; Oct 22 2026 probe says a
team must add a valid payment method before using free credit
(otherwise `403 customer_verification_required`), and buying credit
permanently ends the recurring $5 credit. Treat as **card-likely**.
· **Free quota:** **$5 gateway credit per team per 30 days** (spent at
provider list prices, no markup, BYOK excluded) **plus** a handful of
separately $0-priced routes that do not touch the credit (Oct 2026:
`laguna-s-2.1`, `nemotron-3.5-lightning`, `beam-1.0`; earlier $0 routes
like Ling 3.0 Flash were dropped). At a few dozen short calls/day on
cheap open models, $5/month can last — but a long system prompt on a
larger model burns it faster, and the credit is monthly, not daily.
· **OpenAI-compatible:** Yes, `https://ai-gateway.vercel.sh/v1`.
· **Chinese:** The $0 routes are coding-oriented (Laguna, Nemotron
Lightning, Beam) — usable for short Chinese chat, not chosen for
Chinese quality. The $5 credit could buy Gemini/Flash-class routes,
but that is paid-model spend, not a free tier, and the same Gemini
region caveat applies upstream.
· **Stability / runaway risk:** **Low.** Vercel is a large public-adjacent
company; the gateway is a real product. Risk is the $0 roster turning
over in days and the “buy once, lose the free credit forever” trap.
· **Sources:** https://github.com/mvalentsev/awesome-free-ai-coding/blob/HEAD/providers/vercel-ai-gateway.md (Oct 22 2026)
#### A4. NVIDIA NIM (build.nvidia.com) — USA (vendor host, aggregator-like)
· **Register / card:** NVIDIA Developer Program account + `nvapi-` key;
**no card**. Phone verification in a supported country is reported;
not offered in Russia, Pakistan, Bangladesh, Iran and several other
countries (check the list for the server's country).
· **Free quota:** Models marked “Free Endpoint”: **~40 RPM, reported
10,000 requests/day per account** (one tracker; NVIDIA does not publish
a universal limit — it is account/model-specific and throttling under
load is expected). Signup credits are described as 1,000 inference
credits (up to 5,000 on request) in some guides, but the Nov 2026
AgentDeals record describes it as rate-limited free endpoints with no
per-token billing — treat the credit figure as unverified.
· **OpenAI-compatible:** Yes, `https://integrate.api.nvidia.com/v1`.
· **Free models (overseas):** Gemma 4 31B, gpt-oss-20b, Nemotron 3 Ultra /
3.5 Lightning / 3.5 Super VL, Llama 3.2 Vision, Muse Glimmer.
`[Chinese-origin models also listed — DeepSeek V4.1 Flash, GLM 5.3 Flash,
MiniMax, Kimi — excluded under the strict rule.]`
· **Chinese:** Gemma/Nemotron verdict as above. **Terms trap:** NVIDIA's
API Trial Terms say free use is for **testing and evaluation only,
not production**, prompts/outputs are recorded to improve products,
and each hosted model has an end-of-life date after which it returns
410 Gone. Good for evaluation, not a production backend key.
· **Stability / runaway risk:** **Low for the company, medium for any
single model** (per-model EOL dates, throttling).
· **Sources:** https://agentdeals.dev/vendor/nvidia-nim (Nov 2 2026);
https://mvalentsev.github.io/awesome-free-ai-coding/providers/nvidia-nim/ (Nov 2 2026 probe)
#### A5. Hugging Face Inference Providers — as an aggregator
See §6 above. As a *gateway* it is Tier A (real company, no markup,
one token, pin a provider with `:provider` suffix or `:cheapest` /
`:fastest`). As a *free source* it is Tier “effectively none”
($0–$0.10/month). Its value here is consolidating keys you already have
(Groq, etc.) behind one OpenAI-compatible endpoint — not new free capacity.
#### B1. AIHubMix — commercial One-API-family relay
· **Register / card:** Signup, no card to start. **Sustained free use is
not zero-barrier:** before the first top-up, all free models share
**10 trial calls (never expire)**; after topping up **≥$1**, free
models are usable at roughly **100 requests/day, 10 RPM, 1M tokens/day**
(some heavier models 33–50/day), reset daily. Free IDs end in `-free`
and must be called explicitly, or a paid model name errors on balance.
· **OpenAI-compatible:** Yes, `https://aihubmix.com/v1`, plus Anthropic
format. 800+ models on the paid catalogue.
· **Free models (overseas):** Nemotron 3 Ultra 550B / Super 120B `-free`
routes are the relevant ones; most of the marketed free roster
(GPT-5.5, Gemini 3, GLM, Kimi, MiniMax, Xiaomi MiMo) is either
closed-vendor or Chinese-origin — the “free flagship” claim is the
relay subsidising inference, not an official free tier, and is exactly
the kind of offer that can be withdrawn without notice.
· **Chinese:** Nemotron routes as above; the Chinese-origin free routes
would be better on Chinese but are excluded here.
· **Stability / runaway risk:** **Medium.** Long-running commercial relay
with documentation and a reputation score ~70/100 in one review, but:
no published SLA, DPA, or independently verifiable uptime; request
content passes through the relay (privacy/compliance risk, upstream
providers may retain under their own terms); upstream key sourcing is
not independently verified; claimed 99.99% availability has no public
data. Suitable as a *third* fallback for non-sensitive traffic, not
for the primary path, and never for prompts containing secrets.
· **Sources:** https://freeairouter.com/s/aihubmix.com ;
https://mvalentsev.github.io/awesome-free-ai-coding/models/glm-5.3-flash/ (Nov 2 2026 probe, AIHubMix limits)
#### B2. Fireworks AI — USA (direct inference, not a free tier)
· **Register / card:** Account without a payment method is possible,
but inference needs credits: **prepaid, $1 free credit reported in
Aug 2026** (Oct 2026 tracker says no signup credit is listed on the
pricing page — treat $1 as unverified/stale). An account with no
payment method or no credits is limited to **10 RPM account-wide**
(6,000 RPM with payment + active credits). The only confirmed free
offer is a **training** promo (Muse Glimmer 30B LoRA SFT free until
30 Nov 2026), not inference.
· **OpenAI-compatible:** Yes, `https://api.fireworks.ai/inference/v1`.
· **Verdict:** Not a free option for this brief. As a low-cost paid
fallback, serverless rates are cheap for overseas models
(gpt-oss-120b $0.15/$0.60 per M, Nemotron 3.5 Lightning $0.05/$0.20),
and context is large — but that is a paid decision, not a free tier.
· **Sources:** http://grabaicredits.com/en/deals/fireworks-ai-free-credits (Oct 17 2026);
https://github.com/robhunter/agentdeals/blob/HEAD/artifacts/free-llm-api-index/README.md (Aug 25 2026: $1 credit)
#### B3. Novita AI — small trial, not a free tier
· **Register / card:** No card to sign up.
· **Free quota:** Reports vary ($0.50 baseline welcome voucher, $1 in one
ranking, $10 promotional credit in another, plus a $100/90-day
*Sandbox* credit that Novita does **not** say applies to Model API
calls). No model is priced at $0 on the pricing table as of the
Aug 2026 re-check (“Novita has no free rows left”). Treat as a
**tiny one-time trial**, not a tier.
· **OpenAI-compatible:** Yes. Models are mostly DeepSeek/Qwen/Llama —
under the strict rule only the Llama routes count, and they are not free.
· **Verdict:** Skip for this brief except as a $0.50–$1 smoke test.
· **Sources:** https://freeairouter.com/s/novita.ai ; https://github.com/xyzs996/free-llm-api/blob/HEAD/README.md (Aug 22 2026)
#### B4. Chutes — decentralised (Bittensor subnet 64), USA/community
· **Register / card:** `cpk_` API key; pay-as-you-go, no subscription,
no minimum.
· **Free quota:** **No free tier at this time** (Oct 2026 comparison:
“We do not offer a free tier”). An earlier “$5 deposit for 200/day”
tier was removed in steps Aug 2025–Mar 2026. **However**, the live
catalog on 4 Nov 2026 lists some models at **$0.0000/M input/output**
(e.g. DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro-0813) — these are
subsidy-dependent, mostly Chinese-origin, and not a published free tier.
· **OpenAI-compatible:** Yes, `https://llm.chutes.ai/v1`.
· **Chinese:** The $0 routes are Chinese-origin (excluded); overseas
routes on Chutes are paid. TEE (confidential compute) is a genuine
differentiator for privacy on paid routes, but it does not make the
$0 routes a stable free tier.
· **Stability / runaway risk:** **High for free use.** Compute comes from
independent miners; availability depends on volunteer/subsidised
capacity, model list and $0 pricing can change with subsidies, and
there is no SLA. Do not use as a primary or sole fallback.
· **Sources:** https://arcanumrpgs.com/blog/openrouter-vs-nanogpt-vs-chutes/ (Oct 11 2026);
https://chutes.ai/llms.txt (Nov 4 2026 catalog)
#### C1. Pollinations — Berlin, open-source
· **Register / card:** **None for the legacy lane.** New
`gen.pollinations.ai` requires a key and Pollen credits ($1 ≈ 1 Pollen;
free tier reported ~1.5 Pollen/week, tightening).
· **Free quota (legacy, verified Nov 2 2026):** Anonymous, per IP,
**one text request at a time, at least 30 seconds apart** (images
45s apart); beyond that, 402 with a Retry-After. Legacy lane serves
**one model: `gpt-oss-20b`** (the old “openai/mistral/llama” roster is
stale). The marketed GPT-5 / Gemini / Grok roster belongs to the
keyed Pollen service, not the free lane.
· **OpenAI-compatible:** Yes, legacy `POST https://text.pollinations.ai/openai`
and new `https://gen.pollinations.ai/v1`.
· **Chinese:** gpt-oss-20b only — same gpt-oss Chinese caveat as Groq,
but on the smaller 20B model, so weaker still. Short chat/translation
only, and expect queueing.
· **Stability / runaway risk:** **High.** Shared public endpoints,
latency fluctuates with load, “no uptime guarantee, limits, models
and behaviour can change”, anonymous lane requires attention to prompt
privacy (no account, no DPA). 30-second spacing also means it cannot
serve even a few dozen *bursty* requests, and it must never be the
path for a long system prompt containing anything sensitive.
· **Sources:** https://mvalentsev.github.io/awesome-free-ai-coding/providers/pollinations/ (Nov 2 2026 probe);
https://raw.githubusercontent.com/pollinations/pollinations/master/APIDOCS.md (Classic API limits)
#### C2. LLM7.io and similar free gateways
· **Register / card:** Free token from dash.llm7.io (account needed for
the token; exact signup friction varies by report).
· **Free quota (reported):** 1 request/second, 60/minute, 250/hour,
**100,000 tokens/24h** (input+output). Provider states quotas “may be
reduced without notice”.
· **OpenAI-compatible:** Yes (gateway-style).
· **Chinese / stability:** Model roster is gateway-dependent and
changes; treat Chinese quality as unverified per-route, and
stability as **high risk** — same Tier C warnings as Pollinations,
plus an explicit “may reduce without notice” clause. Emergency
fallback only.
· **Source:** https://agentdeals.dev/vendor/groq (alternatives table, LLM7 row, Oct 2026)
#### C3. Other names that appear in free lists — why they are not shortlisted
· **Ollama Cloud:** Free plan replaced 26–27 Oct 2026 by $0 pay-as-you-go
(per-token billing, no free tier); local Ollama self-hosting is free
but is not an API service. (Source: https://agentdeals.dev/vendor/ollama-cloud)
· **Cohere:** Appears in “no-card free” roundups, but no verified free
chat tier with usable limits was found in this pass — do not assume it.
· **NanoGPT:** No free API tier (one free model, website only, Oct 2026).
· **OpenCode Zen / FreeInference / Regolo / Opper:** Niche or
client-locked free routes (e.g. Zen free models work only inside
OpenCode, 403 otherwise). Not suitable as a generic server-side backend.
| Dimension | Official free tier (Gemini, Groq, Mistral, Cloudflare) | Established aggregator (OpenRouter, Requesty, Vercel, NVIDIA NIM) | Commercial / community relay (AIHubMix, Chutes $0, Pollinations, LLM7) |
|---|---|---|---|
| Free capacity | Published (or console) per-provider cap; single model family | Often *more* daily headroom (Requesty 200/day) or fallback across providers | Small, per-IP, or subsidy-dependent; can vanish |
| Model choice | Narrow, stable IDs | Wide, but roster churns monthly | Widest on paper, least trustworthy in practice |
| Chinese quality | Best at Gemini; otherwise model-dependent | Same models as official, via another hop | Same models, plus risk of substituted/quantised serving |
| Privacy / training | One party's terms; Gemini free trains, Groq does not (with logging caveats) | Two parties' terms (gateway + upstream); per-route training flags differ | Relay sees the prompt; often no DPA/SLA; anonymous lanes have no recourse |
| Stability / runaway risk | Low–medium (Google/Groq can cut limits; GitHub retired) | Low–medium (company + paid business; no SLA on free) | Medium–high (subsidy, miner, or donation funded; no SLA; 410/402 changes) |
| Cost to escape | Change base_url + key | Same — OpenAI-compatible, so switching is cheap | Same wire format, but model IDs differ (`-free` suffixes, etc.) |
| Best use here | Primary (Gemini) + direct fallback (Groq) | Second key / fallback chain (OpenRouter → Requesty) | Emergency third fallback for non-sensitive traffic only |
**Recommended fallback chain for this backend (overseas only):**
Gemini Flash-Lite → OpenRouter Gemma 4 / Nemotron → Requesty Gemma 4 /
Nemotron → Groq gpt-oss-120b (spaced ≥60s apart for long prompts) →
Cloudflare Workers AI Gemma. Keep all keys server-side, pin exact model
IDs, retry with backoff on 429, and log which provider served each
request so a silent model substitution is visible. Never put the
production system prompt through Tier B/C relays.
---
| Rank | Option (overseas model) | Free shape | Context | Fits long prompt? | Chinese verdict | Main risk |
|---|---|---|---|---|---|---|
| 1 | **Gemini Flash-Lite** (Google) via AI Studio | Permanent, no card*; per-project caps, ~500 RPD reported | 1M | ✅ Yes, by far | **Best overseas** — flagship multilingual, idiomatic ZH, strong ZH↔EN | Region: API list excludes mainland China/HK/Macau; free trains on prompts; caps unpublished/shrinking |
| 2 | **OpenRouter** `google/gemma-4-31b-it:free` or `nvidia/nemotron-3.5-super-vl:free` | Permanent, no card; 20 RPM / 50 RPD (1,000 after $10 lifetime) | 262K / 1M | ✅ Yes | Supported (Gemma 140+ langs; Nemotron lists ZH) but not native-grade; test on own prompts | Roster churn; upstream 429s; per-route training/retention |
| 3 | **Requesty** free Gemma 4 / Nemotron | Permanent, no card; 200 RPD shared | 262K / 1M (per model) | ✅ Yes | Same as OpenRouter | NVIDIA-hosted free routes may train/retain 30 days; roster churn |
| 4 | **Groq** `openai/gpt-oss-120b` | Permanent, no card; 30 RPM / 1,000 RPD / 8K TPM / 200K TPD | 131K | ⚠️ Yes, but space calls (8K TPM) | Usable short chat/basic translation; weak native Chinese (C-Eval <30% in one eval) | TPM is the binding cap for long prompts; model cuts without notice |
| 5 | **Cloudflare Workers AI** Gemma 4 (preferred over Llama here) | Permanent, no card; 10K Neurons/day | 128K | ✅ Yes (input-heavy is cheap in Neurons) | Gemma: supported; Llama 3.3: ZH officially unsupported — do not pick Llama for Chinese-first use | Neuron accounting is opaque; free plan hard-stops at budget |
| 6 | **Mistral** free Nemo/Ministral | Free mode, no card; caps unpublished | 128K | ✅ Yes | Serviceable small-model Chinese; below Gemini | Capable Small/Medium NOT free; $10 credit line already removed |
| 7 | **SambaNova** Gemma 4 preview / gpt-oss-120b | Permanent, no card; 20 RPM / 20 RPD / 200K TPD per model | 128K | ✅ Yes | Gemma preview better than Llama (ZH unsupported); preview may vanish | 20/day is below “a few dozen” on one model |
| — | **NVIDIA NIM** Gemma 4 / Nemotron (evaluation only) | No card; ~40 RPM; trial terms | 262K / 1M | ✅ Yes | As OpenRouter | Trial terms: testing/evaluation only, not production; per-model EOL |
| — | **Vercel AI Gateway** $0 routes / $5 credit | Card-likely; $5/30 days + $0 routes | Varies | ✅ Yes | Coding-oriented $0 routes, not Chinese-chosen | $0 roster turns over in days; buying credit kills free credit |
**Explicitly NOT in the shortlist (and why):**
· GitHub Models — retired 30 July 2026.
· Together — no free tier, $5 minimum prepay.
· Cerebras — $5 trial, card required, expires in 30 days.
· Hugging Face alone — $0–$0.10/month, not a backend tier (use only as BYOK router).
· Fireworks / Novita / Chutes — no usable standing free tier for this use case.
· Pollinations / LLM7 / AIHubMix free routes — emergency fallback only (see Tier B/C).
· Any Chinese-origin model (Qwen, DeepSeek, GLM, Kimi, MiniMax, Xiaomi MiMo)
— excluded by the overseas-model rule, even when hosted overseas.
*If that rule is relaxed, the ranking changes materially: Groq Qwen 3.8
would rank ~2nd on Chinese quality, and OpenRouter/HF GLM routes would
be credible fallbacks.*
**Flags the brief asked for:**
· **Would NOT fit a long system prompt (small input caps):** None of the
ranked overseas options fails on context alone — the smallest ranked
context is 128K, and even Cerebras' trial 65K would fit several thousand
characters. The real “does not fit” failures are rate-based: Groq 8K TPM,
SambaNova 20/day, OpenRouter 50/day (unfunded), and Pollinations'
30-second spacing.
· **Free credits expire quickly:** Cerebras $5/30 days (card required);
Together's old $25 is gone; Novita $0.50–$1 (unverified, tiny);
Fireworks $1 (unverified); Vercel $5/30 days (and buying credit ends it);
NVIDIA NIM per-model EOL dates. None of these is a permanent tier.
· **Long-prompt + Chinese + low-volume, overseas-only, in one line:**
Gemini Flash-Lite first, OpenRouter/Requesty Gemma-Nemotron as the
fallback chain, Groq gpt-oss only if Chinese quality is verified on the
site's own prompts.
Researched in the first pass and deliberately excluded by the follow-up:
智谱 BigModel (GLM-4.7-Flash permanently free, 200K context, phone signup,
1-concurrency — strongest domestic free option, but excluded);
硅基流动 SiliconFlow (small open models free + ~¥14 voucher, 180 days);
火山引擎豆包 Ark (500K tokens/model, “安心体验” mode, pauses at cap);
阿里云百炼 Bailian (typically 1M tokens/model, 90 days, Beijing region only,
real-name required, auto-converts to pay-per-token unless “用完即停” is on);
百度千帆 Qianfan (typically 1M tokens/model, 3 months);
DeepSeek (5M tokens, 30 days, conflicting report of 1M/3 days in 2026);
Kimi/Moonshot (¥15 voucher, recharge ≥¥1 to start, flagship K3 excluded from voucher);
腾讯混元 Hunyuan (1M tokens shared, 1 year, migrating to TokenHub, real-name required);
讯飞 Spark (Lite permanently free, but non-OpenAI protocol for the classic API).
These are recorded here only so the exclusion is auditable; none appears
in the ranked shortlist above.
· Exact current Gemini free-tier RPM/RPD for the account's project and
region (Google no longer publishes them; console is authoritative).
· Whether EEA/UK/CH accounts need billing enabled for Gemini free tier
(sources conflict).
· Whether Hugging Face free accounts currently get $0 or $0.10/month
(Sep vs 21 Oct 2026 sources conflict; either way, unusable as a tier).
· Vercel AI Gateway: card required or not for the $5 credit (Sep vs
Oct 2026 sources conflict).
· Fireworks $1 credit and Novita $0.50/$1/$10 figures (trackers disagree;
all are too small to matter for this brief).
· Requesty RPM (not published in sources found); Pollinations' current
Pollen free allowance (~1.5/week, tightening, unverified on the official page).
· No live API calls were made in this research; all “free model answered”
claims are from third-party probes (notably the 3 Nov 2026 OpenRouter probe).
· The user's server egress region was not specified in the mission;
Gemini's region exclusion (mainland China/HK/Macau) could be decisive
and must be confirmed against the actual deployment region.
· No page in this research instructed the agent to do anything outside
the mission; no such instruction was followed.
Key URLs (verbatim, as returned by search / page fetches). Full per-platform
sources are inline above; the most load-bearing are:
· https://agentdeals.dev/gemini-api-pricing-2026
· https://www.morphllm.com/gemini-api-pricing
· https://www.21cloudbox.com/support/google-gemini-china.html
· https://learn.microsoft.com/lv-lv/azure/foundry/foundry-models/how-to/quickstart-github-models
· https://toolfreebie.com/groq-vs-cerebras-vs-gemini/
· https://klymentiev.com/blog/openrouter-free-tier
· https://www.teamday.ai/blog/best-free-ai-models-openrouter-2026
· https://www.requesty.ai/free-models
· https://docs.requesty.ai/quickstart
· https://www.aifreeapi.com/en/posts/mistral-api-free-tier
· https://toolfreebie.com/mistral-free-api/
· https://www.spheron.network/blog/free-gpu-cloud-credits-2026/
· https://aifree.dev/learn/free-hugging-face-inference
· https://blog.laozhang.ai/en/posts/free-ai-api-tiers-compared
· https://agentdeals.dev/vendor/cerebras
· https://www.logeshwaran.org/2026/10/cloudflare-workers-ai-pricing-free.html
· https://arxiv.org/html/2508.12461v2
· https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf
· https://docs.api.nvidia.com/nim/reference/meta-llama-3_3-70b-instruct
· https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/models/nemotron-3-super-120b-a12b
· https://github.com/mvalentsev/awesome-free-ai-coding/blob/HEAD/providers/vercel-ai-gateway.md
· https://github.com/mvalentsev/awesome-free-ai-coding/blob/HEAD/providers/nvidia-nim.md
· https://github.com/mvalentsev/awesome-free-ai-coding/blob/HEAD/providers/requesty.md
· https://freeairouter.com/s/aihubmix.com
· http://grabaicredits.com/en/deals/fireworks-ai-free-credits
· https://arcanumrpgs.com/blog/openrouter-vs-nanogpt-vs-chutes/
· https://mvalentsev.github.io/awesome-free-ai-coding/providers/pollinations/
· https://raw.githubusercontent.com/pollinations/pollinations/master/APIDOCS.md
· https://agentdeals.dev/vendor/nvidia-nim
· https://agentdeals.dev/vendor/groq
· Official docs verified by direct fetch 4 Nov 2026: https://docs.bigmodel.cn/cn/guide/start/pricing (GLM-4.7-Flash = free, 200K — domestic, excluded),
https://docs.bailian.console.aliyun.com/zh/model-studio/new-free-quota (1M/model, 90 days — domestic, excluded),
https://ark.volcengine.com/region:cn-beijing/docs/free-tokens-only-mode (500K tokens, 安心体验 — domestic, excluded).