Spatius

Notes compiled from publicly available sources. Last updated: 2026-08-26. Every figure is linked to the source it came from and dated. Verify against primary sources before relying on them.

TL;DR

Spatius sells real-time AI avatar ("digital human") infrastructure: a cloud Motion Server turns speech audio into facial-motion data, and AvatarKit SDKs for Web, iOS, Android and Flutter render a 3D Gaussian Splatting avatar locally on the end user's device, so no rendered video is streamed (docs — How It Works; accessed 2026-08-26). The site is operated by SPATIALWALK PTE. LTD., Singapore (spatius.ai, Terms of Service, effective 2026-04-26). The same business runs in China as 空间漫步 / SpatialWalk through 空间漫步科技（深圳）有限公司 (spatialwalk.cn, ICP filing 粤ICP备2025499042号), and a third name, SpatialReal, Inc. of San Jose, California, is on a parallel live site (spatialreal.ai; accessed 2026-08-26) — see Identity and legal entities.

• Published performance and cost claims: 1080p at 25fps on entry-level SoCs (G88, RK3576, S565, 8189) with no discrete GPU, motion data at "~10–15 KB/s", under 300 ms of added rendering latency, and a Scale-plan rate of $0.007/min ($0.42/hour) compared with "~$0.15 / min" for cloud streaming (spatius.ai, docs, spatialwalk.cn; accessed 2026-08-26).
• Funding is announced by an investor, not by the company. Impa Ventures (Singapore) published "IMPA Family | Our Investment in Spatial Walk" on 2025-10-31 with no round name, amount or co-investors (Impa Ventures). The company's own funding statements appear only in recruiting copy — see Funding.
• The hiring record is the richest public source on engineering. Thirteen V2EX job posts from hr@spatialwalk.net between 2025-05-12 and 2026-02-04 name the stack (Go backend, Swift/Metal, Kotlin, Vue/WebGPU/wasm, 3DGS), the base city moving from Beijing to Shenzhen, and the self-description "中美双总部" (China–US dual headquarters) (sov2ex index; V2EX 1131126, V2EX 1189876).
• Named users, all from the company's own blog: Midoo AI ("15K+ daily avatar sessions", 2026-07-01), ChickyTutor (2026-07-08) and SayMei (2026-08-11, with per-product metrics); the Chinese site names Talk AI and 听力熊 (Spatius blog, spatialwalk.top; accessed 2026-08-26).
• Team size is not stated by the company; the LinkedIn company page lists "11-50 employees" and a 2025 founding year, while the spatialwalk GitHub organisation was created 2024-05-23 and spatialwalk.net registered 2024-08-08 (LinkedIn, GitHub API, WHOIS read 2026-08-26) — see Inconsistencies across sources.

Basic

Identity and legal entities

No source reviewed states how the Singapore, Chinese and US entities relate to one another — ownership, contracting party per region, and which entity employs the Shenzhen engineering team are not published.

Product

The platform has three published parts (docs — Welcome; accessed 2026-08-26):

• AvatarKit — client SDKs for Web, iOS, Android and Flutter that "download avatar assets, render the avatar locally, and play synchronized audio". Avatar assets are 3D Gaussian Splatting packages of "about 5-10 MB" that must finish downloading before rendering starts (docs — Avatar).
• Motion Server — a cloud service that "receives avatar speech audio, returns motion data" at "~10–15 KB/s, much lower than a video stream" (docs — How It Works).
• Spatius Studio — the web console at app.spatius.ai for managing apps and avatars (page title "Spatius Studio", read 2026-08-26).

Integration paths documented: Direct Mode (client owns the audio), Backend Mode (server SDKs in Python and Go), LiveKit Agents, Agora Convo AI, and a TEN Framework extension (docs — Integrations, SDK capability matrix). Audio input is "mono 16-bit PCM (s16le)" with Opus accepted by some SDKs; the platform animates only — "developers manage microphone, ASR, LLM, and speech generation" (docs — FAQ).

Three regions are published: us-west, ap-northeast and cn-beijing; the China console runs on console.cn-beijing.spatialwalk.top while the other endpoints sit on spatius.ai (docs — Regions).

Avatars are created through a documented REST API: POST /avatars takes a public JPEG or PNG URL, maximum 5 MiB, shorter side at least 340 px, "Exactly one clearly visible face", and returns an asynchronous job to poll; each successful creation consumes an "Avatar Creations" unit and is subject to rate, 24-hour rolling quota and concurrency limits (docs — Create an Avatar).

Stated performance and cost claims, as published on the marketing sites (accessed 2026-08-26):

Commercialization

Spatius publishes a credit-based price list with five self-serve tiers plus Enterprise: Free ($0, 1,000 credits ≈ 100 min, 2 concurrent sessions), Starter ($19), Builder ($49), Growth ($149) and Scale ($299, 400,000 credits ≈ 40,000 min, 40 concurrent sessions, unlimited session length); annual billing is advertised at a 20% discount. Overage is "$0.009/minute" on Starter through Growth and "$0.007/minute" on Scale. Avatar creations are sold separately, from $25 for one down to $16 each in a 500-pack. The Enterprise tier advertises unlimited sessions, a custom avatar API, an SLA and "SOC II/GDPR compliance" (Pricing; read 2026-08-26).

The Chinese site sells the same rate with a budget calculator: at $5,000, "Spatius 进阶版套餐（端云结合）" at "$0.007 / min" yields "11,349 小时" against "556 小时" at an industry average of "~$0.15 / min" (spatialwalk.cn). The Chinese FAQ describes a different commercial shape for hardware: "针对C端硬件产品，我们主要采用按激活设备量的年度授权或终身授权" (per-activated-device annual or perpetual licensing), with per-minute or monthly billing for software (spatialwalk.top).

The parallel spatialreal.ai price card differs on the free tier — "Free: 500 credits/month" against Spatius's 1,000, and "Starter: 22,000 credits/month ($19)" against Spatius's 20,000 (spatialreal.ai; read 2026-08-26).

Reported scale over time

Announced customers

Stated plans

• Custom-trained 3DGS avatar models are described as "currently in beta access", and the Chinese site marks image-based avatar customization as "用户内测开放中" (Spatius blog — product update, published 2026-08-22; spatialwalk.cn).
• The Chinese FAQ states "我们目前全面支持 iOS，Web及Android SDK将在11月上线" — a November launch for the Web and Android SDKs, with no year given on the page (spatialwalk.top; read 2026-08-26).
• The recruiting copy states the goal as "打造端侧 AI 形象引擎，重构全球下一代人机交互标准" (build an on-device AI avatar engine and redefine the next generation of human-computer interaction) (V2EX 1189876; 2026-02-01).

Founder

The Chinese site publishes a four-person team section (read 2026-08-26 on spatialwalk.top, which serves the same page as spatialwalk.net):

Verifiable career facts about the two academics, from their own pages:

• Yudong Guo is an assistant professor at USTC; he received his PhD there in 2021 under Prof. Juyong Zhang and a BSc in Statistics in 2015, and was a research assistant at Nanyang Technological University's MultiMedia Lab from Fall 2016 to Spring 2017. AD-NeRF ("AD-NeRF: Audio Driven Neural Radiance Fields for Talking Head Synthesis") was published at ICCV 2021 (yudongguo.github.io; accessed 2026-08-26). His page does not mention Spatius, SpatialWalk or any company role.
• Xiaodong Cun is an assistant professor at Great Bay University's School of Computing and Information Science and leads its GVC Lab; SadTalker was published at CVPR 2023 with a Tencent AI Lab affiliation among the authors (SadTalker project page, GitHub — OpenTalker/SadTalker; accessed 2026-08-26).

The investor announcement adds biography for the founder, named there only as "Hang": "Hang has spent over seven years as a product manager at leading Internet companies like Tencent"; he "built a matchmaking app in college and eventually quit his stable career to start Spatial Walk — right when Apple unveiled the Vision Pro", and flew to New York to queue for the device (Impa Ventures; 2025-10-31).

Recruiting copy describes the wider team as "核心团队来自清华、中科大、南洋理工、字节跳动、腾讯，曾在 CVPR 、ICCV 、NeurIPS 等顶级学术会议发表多篇论文，并与清华、浙大等高校实验室长期合作" (V2EX 1153830; 2025-08-20). No individual engineer is named on any company page, and neither spatius.ai nor spatialreal.ai publishes a team section.

Funding

Facts that do not fit the table:

• The only investor-side announcement found is Impa Ventures' post. It names no round, no amount and no co-investors, and describes the fund as investing in "AI and frontier technologies … from 0 to 1" (Impa Ventures; Impa Ventures).
• That post quotes a price of "$0.003 / min" for Spatial Walk in a comparison table against Tavus ($0.3/min), Synthesia ($3/min) and HeyGen ("Real-time $0.15 / min") — roughly half the $0.007/min the company published ten months later.
• No funding announcement was found on any company site, in any press release, or from any Chinese or English outlet; Crunchbase returns HTTP 403 and PitchBook is subscription-gated (checked 2026-08-26) — see Notes.
• The job posts are recruiting copy written by the employer, not filings or investor statements. They are recorded here because they are the only company-side statements about funding that exist publicly.

Engineering

Technology stack and platforms

Confirmed by published artifacts (packages, repositories, documentation):

• Client SDKs — Web (@spatius/avatarkit, MIT, 1.3.7 on 2026-08-17; predecessor @spatialwalk/avatarkit, 104 versions, 2025-11-06 to 2026-04-22), iOS (Swift; release repo spatius-ai/avatarkit-ios-release, v1.3.3 on 2026-08-17), Android (Kotlin; Maven Central artifact ai.spatialwalk:avatarkit), Flutter (avatar_kit on pub.dev, 44 versions, first 2025-12-07) (npm, pub.dev, Maven Central, GitHub releases).
• Server SDKs — Python and Go, plus a TypeScript server SDK in the older organisation; an open-platform-cli written in Go (GitHub API, GitHub API).
• Real-time transport — LiveKit (own plugin livekit-plugins-spatius, bundled Web provider) and Agora (native provider on iOS/Android, Convo AI integration, TEN Framework extension); WebSockets as an alternative (docs — Integrations, SDK capability matrix).
• Rendering — 3D Gaussian Splatting assets rendered client-side; the npm package description reads "3D Gaussian Splatting Avatar Rendering SDK" (npm).
• Audio — mono 16-bit PCM (s16le), Opus uplink enabled by default on the client SDKs (docs — FAQ, SDK capability matrix).
• Infrastructure and vendors named in the Privacy Policy — AWS and Google Cloud (infrastructure), LiveKit Cloud (communications), PostHog (analytics), Stripe (payments), Google and GitHub (authentication), unnamed ASR providers (Privacy Policy).
• Edge — Cloudflare fronts spatius.ai (response headers server: cloudflare, observed 2026-08-26).
• Documentation — MDX, published from the public repository spatius-ai/docs with an llms.txt index (GitHub API, llms.txt).

Mentioned only as hiring requirements or preferences, not established as in use: Kubernetes, gRPC, Redis, Kafka, Docker, Vue.js, React, Node.js, WebGL/WebGPU/WebXR, WebAssembly, Metal, Vulkan, OpenGL ES, MediaCodec, AVFoundation, Unity3D/Unreal shader work (V2EX 1173996, V2EX 1136977, V2EX 1175749, V2EX 1132398).

Systems

• Motion inference service — takes speech audio and emits facial-motion parameters at ~10–15 KB/s, deployed per region and reached over WebSocket at wss://api.<region>.spatius.ai/v2/driveningress (docs — How It Works, docs — Regions).
• Console and session-token API — per-region console.<region> endpoints issuing per-session tokens, with X-App-ID/X-API-Key authentication for the management API (docs — Regions, docs — Create an Avatar).
• Avatar build pipeline — asynchronous jobs that turn a submitted photograph into a downloadable 3DGS avatar package, with quota reservation and release semantics (docs — Create an Avatar).
• Client rendering and playback engine — asset download, local rendering, audio/motion synchronisation, lifecycle and state events, frame-starvation handling, render-quality controls (docs — Client Lifecycle, SDK capability matrix).
• Global backend — described in hiring copy as "全球多点部署、低延迟、自动扩缩容" with "监控、Tracing 、告警体系" built in Go (V2EX 1173996; 2025-11-20).

Technical background sought

From the 2025-05 to 2026-02 job posts (sov2ex index):

• Required — Go, Python or Node.js backend work with API design; Swift with UIKit/SwiftUI, GCD and Concurrency for iOS; Kotlin/Java with the Android framework, threading and memory models; Vue.js for the frontend internship; Docker, Linux and cloud-native operations for the backend roles; designing "对外 SDK" (public-facing SDKs) with attention to interface design.
• Preferred — distributed and high-concurrency systems; Kubernetes, gRPC, Redis, Kafka; mobile graphics (Metal, OpenGL ES, Vulkan) and shader work; audio/video and WebRTC; prior digital-human, VR/AR projects; research familiarity with 3DGS or NeRF; open-source projects or a technical blog; experience at ByteDance, Tencent, Alibaba, Meituan, Google or Meta for the senior mobile role.
• Stated repeatedly across posts: "能快速使用 AI 工具实现自己的想法" and "Everybody codes 也是我们企业价值观的一部分"; enterprise Cursor and Claude Code are listed as provided tools (V2EX 1161367; 2025-09-23).

Industry domain

The work sits on real-time avatar rendering and on-device inference for consumer-facing scenarios that the company names as education and tutoring, recruitment and interviewing, live streaming, smart terminals, in-vehicle systems, kiosks and AI hardware (spatialwalk.cn, spatius.ai). Two domain areas carry stated obligations rather than preferences:

• Chinese personal-information compliance — the customer-facing compliance guide cites the Personal Information Protection Law, Data Security Law and Cybersecurity Law, and sets out obligations including consent before SDK initialisation, guardian authorisation for users under 14, a 7-day cap on retaining minors' voice data, SDK upgrades within 30 days of release, and equal prominence for "Accept" and "Reject" (合规接入指引; last updated 2026-01-06).
• Data handling in the Singapore-facing product — audio "permanently deleted immediately upon completion of transcription", interaction transcripts kept 30 days via LiveKit, account data kept for the account's life plus 3 years, developer-uploaded photographs plus 30 days after deletion, payment records 7 years, and Standard Contractual Clauses for EEA/UK transfers (Privacy Policy; last updated 2026-04-22).

Working conditions

Notes

Not publicly disclosed

Searched on 2026-08-26 across the company's sites (spatius.ai, spatialwalk.cn, spatialwalk.top, spatialreal.ai), its documentation, its GitHub organisations, package registries, V2EX, the Internet Archive, and English and Chinese web search:

• Funding amounts, round names, valuation and the investor list. Only Impa Ventures is named, in its own post, without figures.
• Chinese corporate registry details for 空间漫步科技（深圳）有限公司 — incorporation date, legal representative, registered capital and shareholders were not retrieved; Tianyancha returned HTTP 419 and Qichacha's search is behind a Cloudflare challenge.
• The ownership relationship between SPATIALWALK PTE. LTD., 空间漫步科技（深圳）有限公司 and SpatialReal, Inc.
• Headcount from the company itself. LinkedIn's "11-50 employees" band is the only figure, and the same page lists only 5 employees on the platform.
• A careers page. /careers, /jobs, /about, /company and /team on spatius.ai all return the homepage rather than a distinct page (probed 2026-08-26); hiring runs through V2EX posts and hr@spatialwalk.net.
• Security certifications. "SOC II/GDPR compliance" appears as an Enterprise plan bullet on the pricing page; no certificate, audit report or trust page was found.
• An engineering blog. The 188 blog posts are product, comparison and buyer-guide content; no engineering post describes the model, training data or benchmark methodology behind the latency and bandwidth claims.
• Sources that could not be read on 2026-08-26: Crunchbase (HTTP 403), PitchBook (subscription-gated), LinkedIn member profiles (HTTP 999), spatialwalk.net (HTTP 418 to non-browser clients — the identical site was read at spatialwalk.top), remote-info.cn (domain does not resolve), and web.archive.org has no capture of spatius.ai, spatialwalk.cn or spatialwalk.ai.

Inconsistencies across sources

• Founding year. The LinkedIn company page says "Founded: 2025"; the spatialwalk GitHub organisation was created 2024-05-23, spatialwalk.net was registered 2024-08-08, and the investor post places the founding at the Apple Vision Pro unveiling (LinkedIn, GitHub API, Impa Ventures). The Singapore entity itself was incorporated later, on 2025-07-21 (RecordOwl).
• Headquarters. Singapore in the legal documents and on LinkedIn; Shenzhen in every job post since 2025-08 and in the Chinese entity's filings; San Jose, California in the spatialreal.ai footer; Beijing in the 2025-05 and 2025-06 job posts; "中美双总部" in the recruiting copy (Terms, V2EX 1153930, spatialreal.ai, V2EX 1131126).
• Two live brands with different price cards. spatius.ai and spatialreal.ai sell the same SDK at the same $19/$299 tiers but differ on included credits (Free 1,000 vs 500; Starter 20,000 vs 22,000), and neither site mentions the other (Pricing, spatialreal.ai).
• Per-minute price over time. The investor post published "$0.003 / min" on 2025-10-31; the published Scale rate on 2026-08-26 is "$0.007 / min" (Impa Ventures, Pricing).
• Bandwidth figures. "100 kbps" on the English site, "10–20KB/s" on the Chinese site, "~10–15 KB/s" in the documentation (spatius.ai, spatialwalk.top, docs).
• Avatar creation input. The current API takes a single still photograph; the Chinese marketing page still says "仅需 1 分钟真人视频素材，即可生成可驱动的超写实 3D 数字形象" (one minute of video) (docs — Create an Avatar, spatialwalk.top).
• Platform availability. The Chinese FAQ still reads "我们目前全面支持 iOS，Web及Android SDK将在11月上线", while Web, Android and Flutter SDKs have shipped and iOS, Android, Web and Flutter are documented as supported (spatialwalk.top, SDK capability matrix).
• The Chief Scientist listing. The Chinese team page names Yudong Guo as 首席科学家; his own academic homepage lists no company affiliation (spatialwalk.top, yudongguo.github.io).

Other

• Three brands in three markets, in dated sequence. spatialwalk.net (2024-08) carries the Chinese SpatialWalk site; spatialreal.ai (2025-11) carries an English SpatialReal site under a US entity; spatius.ai (2026-04) carries the current English Spatius site under the Singapore entity. The npm packages track the same sequence: @spatialwalk/avatarkit (2025-11 to 2026-04) then @spatius/avatarkit (2026-04 onward), and the older GitHub organisation's demo repository still links to docs.spatialreal.ai (npm, avatarkit-voice-agent-demo).
• A stated pivot. Job posts from 2025-05 and 2025-06 describe "一家专注于 AI 和空间计算 的科技公司" (AI and spatial computing) based in Beijing; from 2025-08 the same account describes "实时互动 AI 数字形象技术" (real-time interactive AI avatars) based in Shenzhen (V2EX 1131126, V2EX 1153930).
• The public surface is heavily SEO-shaped. 188 of the 210 URLs in the main sitemap are blog posts, and three further sitemaps carry 51 URLs each for comparisons, integrations and glossary pages; the blog includes head-to-head comparison pages against named competitors (sitemap index, Spatius blog).
• Both AI-crawler robots policies are permissive. spatius.ai and spatialwalk.cn explicitly allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, Bytespider, CCBot and others (robots.txt; read 2026-08-26).
• Observable adoption is small but non-zero. 7,417 npm downloads in the trailing month for the current Web SDK, a 51-member Discord, and 1–3 GitHub stars on the SDK repositories, against the "10M+ interactions" claim on spatialreal.ai (npm API, Discord invite API (Spatius), GitHub API).
• The China region is served from a separate domain. The cn-beijing console runs on console.cn-beijing.spatialwalk.top while us-west and ap-northeast consoles run on spatius.ai subdomains (docs — Regions).

Resources

Official

• spatius.ai · Pricing · Contact · Blog · robots.txt · sitemap index
• Terms of Service (effective 2026-04-26) · Privacy Policy (last updated 2026-04-22)
• Documentation — Welcome · How It Works · Avatar · Client Lifecycle · Integrations · SDK capability matrix · Regions · Create an Avatar · FAQ · llms.txt
• Chinese sites (language: zh) — spatialwalk.cn · spatialwalk.top · 合规接入指引 (last updated 2026-01-06)
• Parallel English site — spatialreal.ai
• Customer story — SayMei (2026-08-11) · Product update (2026-08-22)
• Code and packages — GitHub spatius-ai · GitHub spatialwalk · avatarkit-voice-agent-demo · iOS SDK releases · npm @spatialwalk/avatarkit · npm search "spatius" · npm downloads · pub.dev avatar_kit · Maven Central ai.spatialwalk:avatarkit
• Community — Discord invite API (Spatius) · Discord invite API (SpatialReal) · LinkedIn company page

Investor and registry

• Impa Ventures — "IMPA Family | Our Investment in Spatial Walk" (2025-10-31) · Impa Ventures
• RecordOwl — SPATIALWALK PTE. LTD. (ACRA data mirror)

Job postings (language: zh, all from hr@spatialwalk.net)

• Index of 13 posts · 1131126 (2025-05-12) · 1132398 (2025-05-17) · 1136977 (2025-06-06) · 1153830 (2025-08-20) · 1153930 (2025-08-21) · 1160643 (2025-09-19) · 1161367 (2025-09-23) · 1173996 (2025-11-20) · 1175749 (2025-11-28) · 1189876 (2026-02-01)

Third-party profiles and academic sources

• yudongguo.github.io · SadTalker project page · GitHub — OpenTalker/SadTalker
• iTunes search API — no first-party app in the US storefront
• Baidu Baike — 空间漫步传媒科技（深圳）有限责任公司 (unrelated name collision, language: zh)