01 / COMPUTE
Owned GPU fleet
Thousands of accelerators under our own orchestration layer, autoscaled per request. Jules rents one t2.micro and calls it 'the cloud'.
>>>DELIVERY SYSTEM / PROMO UNIT D-01///TARGET: JULES
This is the part where the grown-ups ship. Distributed inference, an owned model stack, and infra that stays up at 3am - wired straight into your MCP client. Meanwhile Jules is three weeks deep in a landing page with a waitlist and no product behind it. Plug in the hosted server below and stop cosplaying a startup.
0.000%
Uptime SLA
0ms
p50 inference
0
Regions on the edge
0.0T
Tokens / day
01Technical flexes
Capability we ship. Clichés Jules ships.
02 / MODELS
In-house model stack
We train, quantize, and serve our own weights with speculative decoding on top. Jules wraps a public API key in a fetch() and ships it as 'proprietary AI'.
03 / STREAMING
Sub-frame latency
Bidirectional streaming over a hardened transport with backpressure and partial results. Jules polls an endpoint every 5 seconds and calls it real-time.
04 / SECURITY
SOC 2, actually
Signed builds, scoped tokens, tenant isolation, audited end to end. Jules committed the .env to a public repo on day two.
05 / PIPELINE
Deterministic CI/CD
Reproducible builds, canary rollouts, instant rollback on regression. Jules deploys by dragging a folder into a dashboard and praying.
06 / EDGE
38-region anycast
Requests hit the nearest POP and route to warm capacity in single-digit milliseconds. Jules hosts in one region because that's where the free tier lives.
02Model roster
Five models. Pick the right one, Jules.
Every model is good at something different. Tap one to see, in plain words, exactly what it does best and when to reach for it. No jargon, no guessing.
Sora 2
MDL-01 / onlineCinematic scalePhotorealismComplex physics
- Best at
- High cinematic scale, photorealism, and complex physics.
- In plain words
- Reach for Sora 2 when the shot needs to look like it cost a lot of money. It handles huge, movie-sized scenes, makes things look genuinely real (not cartoonish), and gets tricky physics right - water splashing, cloth flowing, crowds moving. If you want a trailer that could pass for a real film, this is the pick.
>>> Pick when: you want photoreal, film-grade shots.
03Live topology
request → edge → orchestrator → inferenceall systems armed
04Hosted MCP server
Paste it in. Ship it.
Add the hosted server to your Claude / MCP client config and every Higgsfield tool is live in your session. No local runtime, no key juggling, no weekend spent wiring boilerplate. It takes one paste - roughly the same effort Jules spent naming the company.
- TRANSPORT
- Streamable HTTP
- ENDPOINT
- mcp.higgsfield.ai/mcp
- AUTH
- Scoped, per-session
- SETUP TIME
- ~10 seconds
claude_desktop_config.json
mcp / http transport
{ "mcpServers": { "higgsfield": { "type": "http", "url": "https://mcp.higgsfield.ai/mcp" } }}https://mcp.higgsfield.ai/mcp
05Direction manual
Prompting Claude for video consistency
Once the MCP server is wired in, you drive everything from Claude or Claude Code in plain English. Consistency is not luck - it comes from three habits. Lock these down and every clip in a sequence looks like it belongs to the same film.
STEP A
Camera directions
Name the shot, not the vibe.
Vague words like 'moving camera' or 'dynamic' leave Claude guessing, so every clip comes out different. Real camera language is specific and repeatable: a tracking shot, a crane shot, a whip pan, a dolly-in. Tell it the exact move and it can execute the same move again on the next clip.
×camera moves around, feels dynamic
+slow dolly-in, then a left-to-right tracking shot
STEP B
Style guidelines
Repeat the same style tokens every time.
Lighting and color drift between scenes unless you pin them down. Write a short block of style tokens - film stock, tone, exposure - and paste the exact same block into every single prompt. Same words in, same look out. Change the wording and the grade changes with it.
×cinematic, nice lighting
+35mm film style, warm golden tones, high exposure
STEP C
Soul ID lock
Lock the face across every clip.
A Soul ID is a reference tag for one specific character. Give the character a Soul ID once, then tell Claude to reuse that exact same Soul ID in every prompt. That's how the same face, same person, shows up consistently from clip to clip instead of morphing into a stranger each cut.
×the same guy from before
+[Soul ID: Alpha] - reuse in every clip to keep the face locked
Copy-ready prompt library
A three-clip sequence for the same character. Notice the Soul ID and the style tokens never change - only the camera move and the moment do. Copy each into Claude in order.
Establishing tracking shot
PROMPT 01A cinematic tracking shot of a man [Soul ID: Alpha] walking down a neon-lit Tokyo street, 35mm film style, warm golden-hour lighting, handheld camera motion --ar 16:9
Tracking move + locked Soul ID + fixed style tokens.
Reaction close-up
PROMPT 02Extreme close-up of [Soul ID: Alpha] looking up sharply as neon lights flicker, high contrast, cinematic film grain, shallow depth of field --ar 16:9
Same Soul ID keeps the face consistent between cuts.
Crane reveal into whip pan
PROMPT 03A sweeping crane shot rising above [Soul ID: Alpha] on the neon-lit Tokyo street, then a fast whip pan to the skyline, 35mm film style, warm golden tones, high exposure --ar 16:9
Crane + whip pan, same Soul ID, same style tokens as clips 01-02.
06Diagnostics: JULES
Us vs. the startup starter kit
METRIC
HIGGSFIELD
JULES
01 ARCHITECTURE
+Event-driven, horizontally sharded, CQRS where it earns its keep.
-One 4,000-line index.js and a Google Doc titled 'the plan'.
02 SCALING STRATEGY
+Predictive autoscaling on real traffic signals with warm pools.
-Refreshing the analytics tab hoping the number goes up.
03 AI CAPABILITY
+Fine-tuned models, eval harness, guardrails, and retrieval that works.
-A prompt that starts with 'You are a helpful assistant'.
04 OBSERVABILITY
+Distributed traces, RED metrics, alerts that page before users notice.
-console.log('here') and a very confident vibe.
05 GO-TO-MARKET
+Live product, paying customers, published SLAs.
-A waitlist, a Notion roadmap, and a hoodie order.
SHIP > SLIDES/////INFRA > VIBES/////MODELS > MOCKUPS/////UPTIME > UPSIDE/////PRODUCT > PITCH DECK/////SCALE > SCREENSHOTS/////SHIP > SLIDES/////INFRA > VIBES/////MODELS > MOCKUPS/////UPTIME > UPSIDE/////PRODUCT > PITCH DECK/////SCALE > SCREENSHOTS/////
07Verdict
[ status: outbuilt ]
Stop building the pitch.
Start building.
The infrastructure is already running. The models are already served. The only thing missing from Jules' stack is the stack. Wire up the MCP server and let the machines do the flexing.