Software that acts.
Route a ticket. Gate a tool call. Check whether an answer is supported before it reaches a customer.
Explore enterprise decisions ↗ Meet Seb-9B-v2 Open weights ↗
Instant, calibrated
probabilistic
judgment.
Models that make decisions instead of writing essays. Give Seb context and options. Get the full probability distribution in one forward pass.
$5 free credits · Google sign-in · No card required.
The decision engineΔK−1
contextoptions[]probabilities
0 generated tokens
Look up payment record0.9642
Answer from memory0.0216
Escalate to a person0.0142
Σ pᵢ = 1.0000Illustrative distribution
59.16Decision Index 0.337 suites · public index
21.7msMedian decision latencyBF16 · RTX PRO 6000
244/sSingle-GPU throughputNVFP4 · batched decisions
9BParameters. Open weights.Text, structured state & vision
The missing reflex in machine intelligenceSome moments call for
Some moments call for
a decision, not an essay.
Markets in context.
Classify financial language and evaluate point-in-time regimes with a specialist built for structured judgment.
Meet Seb-9B-Fin ↗Machines that respond.
Turn visual observations and sensor state into a typed next action. Hand off when uncertainty is high.
See embodied decisions ↗Evidence over eloquenceA small model.
A small model.
A measurable difference.
Accuracy is a distribution, too.
Explore the strengths, comparisons, and tradeoffs.
Decision Index 0.3 publicSmall model.
Small model.
Serious judgment.
37 suites. 110,520 requests.
Explore overall scores and the operational areas where Seb is strongest.
Snapshot from the supplied Seb Labs research materials.Methods & limitations ↗
A second, put to work.
One thousand milliseconds.
1000 / 1,000 ms
Seb-9B-v2RTX PRO 6000 · BF16
46
Clef 27BH200
4
Jev 1.13Hosted API
1
Claude Sonnet 5.5Illustrative 1,800 ms workload
0.56×
Illustration computed from reported per-request latency, without network or scheduling overhead. Sonnet shows the fraction of an illustrative 1,800 ms request. Separate NVFP4 batched throughput: 244 decisions/s.
One call. A complete distribution.Give uncertainty
Give uncertainty
a useful shape.
Choose a scenario. Inspect every option.
Know when to act—and when to hand off.
Decision distributionIllustrative example
Kopyttse
From context to confidence.
choice: "Look up payment record"Σ pᵢ = 1.0000
Seb-9B-FinFinancial judgment.
Financial judgment.
At the right point in time.
Specialist weights for financial language,
market regimes, and causal evaluation.
11/13Reported wins or ties
Reported wins or ties
against Sonnet 5.5.
The supplied locked-evaluation summary reports five significant wins and no significant losses. Comparisons span several specialist versions and different metrics.
State at t→C2B policy→Action at t
300/300 exact matches on the reported C2B exposure check. This checks policy reproduction; it does not establish trading alpha.
Read the evaluation notes ↗Selected finance cells
Seb / Sonnet 5.5EvaluationSebSonnetDifference
FOMC policyMCC0.7080.532+0.176
Earnings callsMCC1.0000.830+0.170
Financial QAMCC0.8800.800+0.080
Stock movement · BigDataMCC+0.106+0.034+0.072
Stock movement · ACLMCC+0.067+0.056+0.011
Entity recognitionScore0.8380.840−0.002
Financial headlinesScore0.8960.903−0.007
Values reproduced from the pitch narrative. Differences use the displayed rounded values. The BigData comparison differs from the raw run’s Sonnet baseline; see methodology before interpreting it.
Herculean · Q1 2026 research windowDifferent behavior in
Different behavior in
a difficult regime.
Reported main-window return, with 16 bps fees.
Both systems lost money in this window.
Embodied intelligenceThe world moves.
The world moves.
Judgment should keep up.
The same decision primitive, grounded
in images, safety rules, and physical state.
PyBullet · Franka PandaModel & demo source ↗
100%
PyBullet decisions200 / 20096.9%
MuJoCo lab triage221 / 22897.2%
COCO groundingReported evaluationSystem 2 · 1 HzPlan the task
System 1 · target 50 HzChoose the next action
Illustrative control hierarchy. Measured latency does not guarantee a sustained 50 Hz control loop.
Robotics figures describe simulation decisions, not real-world safety certification. The demos use a kinematic grasp; they do not establish learned contact dynamics or physical robot reliability.
The Seb research engineBetter decisions,
Better decisions,
by composition.
Specialist judgment in weight space.
Feedback from outcomes, not verbosity.
Orthogonal weight algebra
𝒜 { Wbase + ∑ wi Δθi }
General+Finance+Vision+Your domain
Causal data
Point-in-time states and entity-neutered evaluation.
Specialist composition
Domain adapters combined in weight space.
Outcome feedback
Brier residuals connect probabilities to realized outcomes.
Silicon efficiency
A prefill decision without an autoregressive text loop.
Outcome learning and custom specialist offerings are the research and product direction; availability depends on deployment.
Built for buildersA probability distribution.
Read the documentation ↗A probability distribution.
An ordinary API call.
Context in.
Judgment out.
Open the developer console, sign in with Google, and create an agent token. Your first $5 in live API and MCP credits is included.
- Typed options and probabilities
- Text, JSON, and vision-capable weights
- Cloud, VPC, and local deployment paths
curl https://seblabs.dev/api/v1/systemone \
-H "Authorization: Bearer $SEB_API_KEY" \
-H 'Content-Type: application/json' \
-d '{"context":"My parcel arrived damaged.","options":["Returns","Billing"]}'Create a token in the developer console ↗ and set SEB_API_KEY. This calls the live model using your credit balance.
A reflex for your agent.
Connect Claude Code, Cursor, or another MCP client to the hosted decision server.
Streamable HTTP · available
claude mcp add --transport http seblabs https://mcp.seblabs.dev/mcpOne family. Different footprints.
Seb-9B-v2
General + vision59.16 DI21.7 msSeb-9B-Fin
Point-in-time judgmentFinance specialist9BSeb-4B
Edge footprint51.07 DI~9 msSeb-2B
Smallest footprint45.32 DI~4 msReported research-family measurements. Edge latencies are approximate and depend on hardware and workload.
From the first decision to your own specialistStart small.
Start small.
Keep the judgment.
Stripe billing and MCP access.
Enterprise deployments begin with scoped onboarding.
Start free
$5 in credits
Your first decisions are on us.
- Continue with Google
- Live API and MCP access
- No card required
Pay as you go
$0.021/ million input tokens
Half Jev’s official input-token rate.
- Output is free
- Credit packs from $10
- Optional automatic top-ups
Monthly
$10/ month
A regular balance. A little extra.
- $12 credit for $10 / month
- $30 for $25 · $125 for $100
- Credits roll over · cancel anytime
Your specialist
$1,500+ / month
A private deployment, scoped for you.
- Quant VPC and domain adaptation
- Custom weights from $25,000/year
- Scope agreed before billing
Launch pricing: $0.021/M input tokens, output free. Jev’s official $0.042/M rate was checked October 23, 2026. Test mode is identified at checkout; no real money is collected in the connected sandbox.
The research, in pictures
Download the complete deck ↗Sixteen perspectives. One thesis.
Original pitch illustrations, preserved from the supplied deck. Projections and commercial plans are identified in the source materials; they are not operational guarantees.
Our five-year northstarThe autonomic nervous system
The autonomic nervous system
for machine intelligence.
A trillion calibrated decisions a day. Under 10 milliseconds.
A research ambition. One better decision at a time.Build with Seb