Free-inference capacity finders — Asia scope, 2026-10-21

Written: 2026-10-21 by worker W102 (finder, Asia providers), model deepseek/deepseek-flash. Scope: the twelve Asia-scope platforms named in the dispatch brief. For each one: what free inference is actually offered (models, quota, expiry), the signup URL, the form fields, and the human-verification step — with a phone number called out wherever it is mandatory. A verdict per platform, ranked, at the end. Method: SearXNG on loopback plus direct fetches of the provider’s own documentation. No fleet host was used to scrape anything; every request originated from this workstation.

This is a record of what the providers’ pages said when read, not a live status. Where a page is JavaScript-rendered and the value could not be read from the page itself, the row says so and the value is marked derived. Nothing here is a credential, and no credential value is reproduced.

0. Summary

#	Platform	Free inference	Gate on signup	Verdict
1	Zhipu / bigmodel.cn	8 free models incl. GLM-4.7-Flash 200K ctx, free vision/image/video	SMS phone (overseas numbers accepted)	pursue — best of set
2	SenseTime SenseNova	Token Plan public beta, 1,500 requests / 5 h, OpenAI-compatible	SMS phone	pursue — highest volume
3	Aliyun Bailian (Model Studio)	1M tokens per model × 70+ models, 90 days	mobile + account info, no card	pursue — broadest catalogue
4	Tencent Hunyuan	hunyuan-lite ¥0, plus one-off new-user pack	mainland real-name (ID)	pursue-with-verification
5	SiliconFlow	permanently-free small models + embeddings; claimed ¥14 credit	phone + ID verification for any free model	pursue-with-caution
6	Baidu Qianfan	ERNIE-Speed/Lite-8K long-term free; new-user grants	phone + ID + face real-name	weak — gated
7	iFlytek Spark	Spark Lite free (8K), 2 QPS	phone + real-name (ID, manual review abroad)	weak — gated
8	StepFun	no free text model; audio previews only	Alipay face scan even for free tier	decline
9	Yi / 01.AI	platform closed 2026-09-03	—	dead
10	Moonshot / Kimi	none — $1 minimum top-up	—	skip (key held)
11	DeepSeek	none on the official platform	—	skip (key held)
12	MiniMax intl	free tier exists, rate-limited	—	skip (ollama-cloud held)

Top three by usable free capacity per unit of onboarding friction: Zhipu, SenseTime, Aliyun Bailian. All three accept a non-mainland phone number, take no card, and speak OpenAI-compatible HTTP.

1. Zhipu — bigmodel.cn open platform (智谱开放平台)

• Signup: https://bigmodel.cn/ (console: https://bigmodel.cn/console/overview)
• Free inference: the model overview page marks eight models 免费 (docs.bigmodel.cn model overview, read 2026-10-21):
  • glm-4.7-flash — free text model, 200K context, 128K max output; agentic-coding tuned (model page).
  • glm-4.5-flash — free, 128K ctx; retiring 2026-01-30, after which requests are auto-routed to glm-4.7-flash. The retirement date has already passed, so treat 4.5-Flash as an alias.
  • glm-4-flash-250414 — free, 128K ctx / 16K out.
  • glm-4.6v-flash, glm-4.1v-thinking-flash, glm-4v-flash — free vision.
  • cogview-3-flash (image), cogvideox-flash (video).
• Quota and expiry: the model pages say “完全免费开放使用” with no token quota and no expiry printed. Rate limits are per account tier and are not on the public model page — unknown, resolve by reading the console after signup.
• Fields: phone number, SMS code, password; optional WeChat QR login.
• Human verification: phone SMS — required. The registration FAQ states overseas numbers are supported (choose the country code), so no mainland SIM is needed. Real-name (实名认证) is offered but is not required to create a key; enterprise real-name unlocks enterprise billing.
• Integration: OpenAI-style POST https://open.bigmodel.cn/api/paas/v4/chat/completions, thinking.type toggles deep-thinking, tool calling and MCP supported.
• Risk: the free list churns (4.5-Flash was retired inside six months). Pin glm-4.7-flash and re-check the free page quarterly.

2. SenseTime — SenseNova Token Plan (商汤日日新)

• Signup: https://platform.sensenova.cn (product page: https://www.sensenova.cn/token-plan)
• Free inference: Token Plan public beta, free of charge — 1,500 requests per 5 hours, stated both in the vendor release and on the product page. Models in the plan include SenseNova 6.7 Flash-Lite (multimodal, document/image understanding) and U1.5 Lite (image generation), and the plan’s OpenAI-compatible base URL is https://token.sensenova.cn/v1. A secondary source lists Kimi K3, DeepSeek V4 and GLM-5.2 as also reachable under the same plan — derived, not verified on a vendor page.
• Quota and expiry: 1,500 requests / 5 h is a rolling window, not a token budget. The word “public beta” is the expiry risk: the free window closes when the beta ends, on the vendor’s schedule and with no notice period published.
• Fields: phone number, SMS code (per a step-by-step integration write-up); no card, no top-up. Whether a non-mainland number is accepted is unknown — resolve at signup.
• Human verification: phone SMS — required; no real-name ID step was described in the sources read.
• Why it ranks above Aliyun: request-count quota, not token quota, and it fronts third-party frontier models. For chat-shaped agent traffic 1,500 calls per 5 h is a bigger practical budget than 1M tokens once per account.

3. Aliyun Bailian / DashScope — Model Studio (阿里云百炼)

• Signup: https://modelstudio.console.alibabacloud.com/ap-southeast-1 (international/Singapore console). Mainland console: https://bailian.console.aliyun.com
• Free inference: first activation issues 1M tokens per model, each model independent, covering 70+ models (Qwen series, DeepSeek, GLM, Kimi, Yi and others). Observed on the official zh page (updated 2026-10-16):
  • Expiry: 90 days from activation, model release or approval, whichever is later. No extension, no re-issue, no reset. Unused quota lapses.
  • Quota is shared between the main account and RAM sub-accounts but not between models; a snapshot name and its undated alias count as two models, each with its own quota.
  • Real-time inference only. Batch, fine-tuning, deployment and OSS are excluded.
  • “Free-quota-exhausted stop” (免费额度用完即停) is a per-model switch, off by default; when on, an exhausted model returns HTTP 403 with AllocationQuota.FreeTierOnly instead of billing.
• Region caveat — conflict, flagged: the zh page read 2026-10-16 says only the Singapore region carries free quota. The English page still says China (Beijing) and Singapore are both eligible. Resolve against the console’s free quota page after signup; do not plan on Beijing.
• Fields: Alibaba Cloud account; profile completion (mobile number + basic info) is mandatory before the platform can be activated. No card for the free quota.
• Human verification: mobile number — required (international numbers accepted on the .com site). ID verification is only needed to continue on pay-as-you-go after the quota is gone.
• Risk: one-shot per account, 90-day clock starts at activation whether or not the quota is used. Activate deliberately, not speculatively.

4. Tencent Hunyuan (腾讯混元)

• Signup: https://cloud.tencent.com/product/hunyuan → console (international: https://www.tencentcloud.com)
• Free inference: hunyuan-lite is the ¥0 tier — input and output priced at zero, 256K context, positioned by Tencent as the free entry tier. New API users additionally get a one-off 1M token pack (community-tracked as running to 2026-12-31); an earlier promotion granted a 100M-token resource pack valid one year. Both figures are derived — the official pricing page did not render server-side in this pass (see claim register).
• Quota and expiry: lite is standing (no expiry published); the new-user pack is one-off and time-boxed.
• Fields: Tencent Cloud account (email/phone), then service activation.
• Human verification: mainland real-name (ID card) is required to enable the model service — this is the blocker for a non-mainland operator. The international site (tencentcloud.com) has a separate onboarding path; whether it offers the same free tier is unknown.
• Verdict: pursue-with-verification. The model is weak, but a standing ¥0 tier is worth a cheap bulk lane if the international path avoids the ID gate.

5. SiliconFlow (硅基流动)

• Signup: https://cloud.siliconflow.cn (mainland) / https://siliconflow.com (global)
• Free inference: the public pricing page (read 2026-10-21) lists these rows as 免费 with no expiry: tencent/Hunyuan-MT-7B (translation), XingChenAGI/Xing4.0-29B, PaddlePaddle/PaddleOCR-VL-1.5, and the embeddings/rerankers BAAI/bge-m3, bge-reranker-v2-m3, bge-large-zh-v1.5, bge-large-en-v1.5. A “new user gets ¥14 credit” claim circulates widely but does not appear in the vendor’s own pricing or agreement pages — treat as unverified.
• Quota and expiry: free rows show no quota; the credit, if it exists, is one-off.
• Fields: phone number + SMS; optional email/WeChat binding.
• Human verification: phone — required, and ID verification is required to use any free model (vendor tutorial). Overseas numbers work at signup, but foreign-origin models require a mainland ID, which an overseas operator cannot supply. Global site behaviour is unknown.
• Verdict: pursue-with-caution. The genuinely free set is small models and embeddings — useful as an embeddings lane, not as chat capacity.

6. Baidu Qianfan (百度千帆 / ERNIE)

• Signup: https://console.bce.baidu.com/qianfan/overview
• Free inference: ERNIE-Speed-8K, ERNIE-Lite-8K and ERNIE-Tiny are long-term free for real-named accounts (7 preset services); new users also get 1M tokens per model, roughly 3 months (community-tracked — the official free-quota page is JavaScript-rendered and returned navigation only in this pass, so the numbers are derived).
• Quota and expiry: lite models standing; grants time-boxed; QPS on the shared free resource pool is not guaranteed.
• Fields: Baidu account, phone, then real-name.
• Human verification: phone + real-name ID + face liveness — required. Baidu gates model service activation on real-name completion; HK/Macau/Taiwan and overseas applicants go to manual review with non-mainland ID documents.
• Verdict: weak. The free models are small and QPS-throttled, and the face gate is the heaviest in this set. Worth one attempt only if an ID document is already available for another platform.

7. iFlytek Spark (讯飞星火 / xfyun)

• Signup: https://www.xfyun.cn/ → console → create an app
• Free inference: Spark Lite is free — the official HTTP API doc states the Lite tier “具有更高的响应速度，支持免费使用”; 8K context, domain=lite. OpenAI-compatible endpoint https://spark-api-open.xf-yun.com/v1/chat/completions exists alongside the native WebSocket API. A separate, larger Spark Pro grant (millions of tokens per month) is a recurring campaign, not a standing tier.
• Quota and expiry: Lite is standing but 2 QPS (community-measured) and the giveaway packs are monthly-reset campaigns.
• Fields: phone + SMS, app name and description, then real-name.
• Human verification: phone + real-name — required. xfyun’s own authentication doc says HK/Macau/Taiwan and overseas users submit valid ID documents for manual review; mainland users get automatic approval.
• Verdict: weak. Free tier is real and permanent, but the model is the smallest in this set and the review step is slow.

8. StepFun (阶跃星辰)

• Signup: https://platform.stepfun.com
• Free inference: none worth onboarding for. The official price table (read 2026-10-21) prices step-3.7-flash at ¥1.35/¥8.1 per 1M tokens and step-3.5-flash at ¥0.7/¥2.1; the only 限时免费 rows are audio previews (stepaudio-3-realtime-preview, -chat-preview, -gen-preview, -music-preview). The old free step-1-flash tier is gone. Step Plan is a subscription from ¥49/month for 400M credit.
• Human verification: Alipay face-scan real-name — required even for the free tier. The FAQ states users on the free plan must complete real-name verification to keep using it, and personal real-name runs through Alipay face recognition. A single real identity covers at most two accounts.
• Verdict: decline. No free text capacity, and the heaviest identity gate in the set.

9. Yi / 01.AI (零一万物)

• Status: the open platform is closed. The 2026-08-10 announcement stopped new registrations and top-ups immediately, and API calls stopped on 2026-09-03; the refund window runs to 2026-12-03.
• Verdict: dead — decline. Yi models remain reachable as a catalogue item inside Aliyun Bailian, which is covered by row 3.

10. Moonshot / Kimi

• Free inference: none. The pricing docs describe pay-as-you-go only and explicitly rule out a subscription; third-party reviews agree there is no permanent free tier and that a $1 minimum top-up is required before the API answers, with a $5 voucher after $5 cumulative top-up. The older “¥15 signup credit” is not in the current docs.
• Verdict: skip. The estate already holds a Moonshot key; a second account buys no free capacity, only another paid balance.

11. DeepSeek

• Free inference: none on the official platform. The pricing and rate-limit pages publish no free tier or trial amount; billing deducts from a topped-up balance first, then granted balance. Aggregator posts claiming a “5M tokens for 30 days” signup grant contradict the official pages and are treated as false here.
• Verdict: skip — the estate holds a key. Free-ish DeepSeek V4 access is available through the SenseTime beta (row 2), which is the cheaper route to the same model family.

12. MiniMax international

• Free inference: a free plan exists — the rate-limit doc lists a 免费用户 tier (e.g. 20 RPM / 1M TPM for the M3.1-Flash class) alongside higher paying tiers. M3.1 Flash is additionally free across plans during its launch window and free on Ollama Cloud until 2026-10-22 23:59 UTC.
• Verdict: skip. Ollama Cloud is already held and covers this model while the promo runs. Revisit after 2026-10-22 if MiniMax’s own free tier outlives the promo.

Ranking and next steps

1. SenseTime — open the account first: highest request volume, OpenAI compatible, no ID step described. Cost: one phone number, minutes.
2. Zhipu — open second: the only provider with a standing 200K-context free text model plus free vision/image/video, and overseas phone signup is documented. This is the one to put in the free-first rotation.
3. Aliyun Bailian — activate last and deliberately, when a 90-day window is actually needed, because the clock starts on activation and never resets.
4. Tencent / SiliconFlow — verify the international onboarding path before spending a phone number.
5. Baidu / iFlytek — only if an ID document is already being submitted elsewhere; the face/review gate is not worth it for Lite-class models.
6. StepFun / Yi / Moonshot / DeepSeek / MiniMax — decline or skip, as above.

Rules of engagement for whoever onboards these. One account per provider per operator. Do not farm quotas with multiple accounts behind one identity: every provider here ties real-name to an identity, several cap accounts per identity, and multi-account quota farming is a terms violation that would put the whole estate’s access at risk. Phone numbers are the scarce resource in this set — a single number can serve several providers, so track which number is bound where before opening the next account.

Claim register

Claim	Class	Evidence
Zhipu lists eight free models incl. GLM-4.7-Flash 200K/128K	observed	docs.bigmodel.cn model overview + model pages, read 2026-10-21
Zhipu accepts overseas phone numbers	observed	registration-faq page, read 2026-10-21
GLM-4.5-Flash retired 2026-01-30 and routes to 4.7-Flash	observed	GLM-4.5-Flash model page notice
Aliyun: 1M tokens/model, 90 days, no reset, real-time only	observed	alibabacloud help zh “新人免费额度”, updated 2026-10-16
Aliyun free quota is Singapore-only	observed, conflicting	zh page says Singapore only; EN page says Beijing + Singapore — unresolved
SenseNova beta: 1,500 requests / 5 h, free	observed	sensetime.com news release + sensenova.cn/token-plan
SenseNova also fronts Kimi K3 / DeepSeek V4 / GLM-5.2	derived	secondary integration write-ups only
SenseNova signup is phone + SMS	derived	integration write-up; not read on a vendor page
StepFun free tier requires Alipay face real-name	observed	stepfun docs FAQ “不进行实名认证会影响使用吗”
StepFun has no free text model	observed	stepfun docs API price table, 2026-10-21
SiliconFlow free rows and ID-verification requirement	observed (pricing) / derived (ID gate)	siliconflow.cn/pricing; vendor tutorial for the ID gate
Baidu free models and grant sizes	derived	official free-quota page is JS-rendered; values from community trackers
Baidu requires real-name + face	derived	vendor onboarding docs and third-party guides
Tencent hunyuan-lite is ¥0 and new users get 1M tokens	derived	official pricing page did not render; aggregator + vendor developer article
iFlytek Spark Lite is free, 8K, 2 QPS	observed (free) / derived (QPS)	xfyun HTTP API doc; QPS from community reports
Yi platform stopped API on 2026-09-03	observed	vendor announcement via two outlets
Moonshot has no free tier, $1 minimum top-up	observed	platform.moonshot.cn pricing docs + two independent reviews
DeepSeek has no official free tier	observed	DeepSeek billing/rate-limit pages; aggregator claims rejected
MiniMax free-plan rate limits	observed	platform.minimax.cn rate-limit doc

What this document does not establish

• Whether any of these providers will serve a UK/EU card-less account without a mainland number, for the ones whose signup pages we could not read end to end.
• Real observed throughput, latency or model quality from the estate. Nothing here was called with a key; this is a paperwork dossier, not a benchmark.
• Whether the free tiers survive the next quarter. Three of the twelve changed or closed inside the last six months (Yi closed, StepFun’s free text tier vanished, GLM-4.5-Flash retired).