Seb Labs developer reference

Context to judgment.

The website exposes a demonstration endpoint and an optional authenticated inference proxy. Demo responses and live responses are explicitly distinguished.

Try a demonstration

Send a JSON body to POST /api/v1/systemone. The four presets return illustrative distributions without calling a model.

curl "$SEB_LABS_URL/api/v1/systemone" \
  -H 'Content-Type: application/json' \
  -d '{"preset":"enterprise"}'

Set SEB_LABS_URL to this website’s origin. Preset IDs are enterprise, finance, robotics, and handoff.

{
  "mode": "illustrative",
  "model": "Seb-9B-v2",
  "state": 1,
  "choice": "Look up payment record",
  "options": [
    { "label": "Look up payment record", "probability": 0.9642 },
    { "label": "Answer from memory", "probability": 0.0216 },
    { "label": "Escalate to a person", "probability": 0.0142 }
  ]
}

These probabilities explain the product interface. They are not model outputs, calibration evidence, or measured latency.

Connect live inference

Live custom context is served by the separately deployed Seb-9B-v2-NVFP4 runtime on ironbccllm GPU1, behind an authenticated HTTPS gateway. The website itself does not host a GPU runtime. Configure these server-only environment variables in Vercel, then redeploy:

SYSTEMONE_BACKEND_URL: the full HTTPS URL of your backend’s /v1/systemone endpoint.
SYSTEMONE_BACKEND_TOKEN: the token accepted by your inference gateway.
SYSTEMONE_API_KEY: an optional owner key for operator verification. Customer keys are provisioned from Google accounts and stored as hashes.

The model contract is state plus a typed questions.decision choice. The website maps your option labels to stable option keys and returns a validated probability distribution.

curl "$SEB_LABS_URL/api/v1/systemone" \
  -H "Authorization: Bearer $SEB_LABS_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "context": "The customer wants to replace a damaged order.",
    "options": ["Returns", "Billing", "Technical support"]
  }'

Limits: 32 KB request body, 12,000 context characters, 2–20 unique options, and 80 characters per option. The upstream deadline is 15 seconds. Missing configuration returns 503 BACKEND_NOT_CONFIGURED; invalid keys return 401; invalid input returns 400. Upstream errors are sanitized. API responses are not cached.

Client keys are not persisted by the playground. A live response includes mode: "live" and measured proxy round-trip latency, which includes network and serving overhead.

Run the open weights

Use the model’s maintained instructions for serving with vLLM, the System One adapter, or Apple MLX. Model licensing and hardware requirements are described in the model card.

Seb-9B-v2 model card and serving guide

hf download ironbcc/Seb-9B-v2 --local-dir seb-9b-v2

# After starting vLLM as described in the model card:
python serve/system1_vllm_backend.py \
  --vllm-url http://127.0.0.1:8000 \
  --served-model seb \
  --tokenizer ./seb-9b-v2 \
  --port 8020 --model-name seb-9b-v2

Expose the adapter through an authenticated HTTPS gateway before connecting Vercel. The loopback addresses above are local serving examples.

Connect with MCP

The stateless Streamable HTTP endpoint is https://mcp.seblabs.dev/mcp. It uses the official MCP SDK. The free illustrative tool works without a key; live calls require an account key and sufficient credits.

claude mcp add --transport http seblabs https://mcp.seblabs.dev/mcp --header "Authorization: Bearer $SEB_API_KEY"

For stdio clients, use the downloadable connector package. It is distributed directly from this website; no npm registry publication is required.

SEB_API_KEY=your_key npx -y https://seblabs.dev/downloads/seblabs-mcp-1.0.2.tgz

seb_demo: inspect one of the four illustrative presets.
seb_decide: submit a text state and 1–10 typed questions (choice, noul, score).
seb_route: classify context among 2–20 caller-provided routing options.
seb_fin_analyze: use the dedicated financial backend, when enabled for your plan and deployment.
seb_batch_decide: process up to 100 states with shared questions and up to four concurrent requests; each item reports its own error or result.

Tools do not execute trades, actuate robots, or call downstream routing targets. The general model is never relabeled as a finance specialist.

A balance for your decisions

Sign in with Google at Account & billing and receive $5 in starter credit once per Google identity. No card is required. The console refreshes usage every five seconds while visible, with 7/30-day charts, per-agent totals and the latest 20 request records. Daily detail begins when tracking is enabled; lifetime totals include earlier usage. Recent requests and agent totals include server inference time and API E2E time, with overall averages for timed requests. API E2E runs from API entry through credit settlement; it excludes the agent-to-API network leg and background telemetry. Older untimed requests are excluded from averages. No prompts are stored. Use Connect to agent to create a separate inference token and copy exact MCP connection instructions, including the secret header, Claude Code command, and Cursor configuration. Or create named auth tokens manually. Tokens share the account balance, are shown once, and can be revoked independently. They cannot access billing or create other tokens.

Launch pricing is $0.021 per million input tokens, with output free: half the official Jev rate of $0.042/M checked October 23, 2026. We meter the model-tokenized prompts, including template, state, instructions and criteria. The current runtime evaluates questions separately, so shared state is counted for each question. Responses include usage.input_tokens and integer usage.cost_nanos (one billion nanodollars = $1). Failed or invalid inference is not charged.

Buy $10, $25, $50 or $100 credit packs. Optional monthly plans add $12 for $10/month, $30 for $25/month, or $125 for $100/month. Credits roll over. Cancel future renewal in Manage billing. Refunds and disputes revoke the corresponding credits.

Connect a card through Stripe, then explicitly enable automatic top-ups. Set a balance threshold, purchase amount, and monthly automatic-charge cap. Checks run on API use or when settings are saved. A low balance can trigger a charge immediately after enabling. Ambiguous payments retain one idempotent attempt; charges requiring authentication pause top-ups. Manual purchases and subscriptions are outside the automatic-charge cap.

The account limit is 120 questions per minute and eight in-flight requests. Batches report per-item errors. Private storage uses conditional writes for reservations and payment deduplication. Stranded reservations are recovered after five minutes. Keys are hashed and displayed once; rotation revokes the old key.

Quant VPC and custom weights require scoped onboarding. The connected Stripe sandbox collects no real money. Live payment activation requires a live merchant account and matching live prices and webhook configuration.

Benchmarks and limitations

The site presents the October 2026 research snapshot supplied with the Seb Labs handoff. It does not rerun these benchmarks or independently certify their results. The original pitch deck and design specification are preserved.

Decision Index 0.3: 37 suites and 110,520 requests, with Seb-9B-v2 at 59.16 overall. Clef 27B scores 61.71 overall, while Seb leads or ties the selected four operational areas. The deck’s “all areas” language is narrower than the full index.
The comparison includes Jev 1.13 at 57.96. Jev 1.14 is reported at 60.70; the site does not imply a win against that later version.
BF16 latency: 21.7 ms p50 and 36.1 ms p95. NVFP4: 21.9 ms p50, 25.2 ms p95, and 244 batched decisions/s. The handoff sometimes labels 21.7 ms as NVFP4; the pitch’s explicit precision breakdown is used here.
Latency comparisons use different hardware, API environments, and workload settings. The impulse race is an arithmetic illustration, not a live or same-hardware benchmark. Throughput and serial latency are separate measurements.
Robotics results are simulation decisions: 200/200 PyBullet and 221/228 MuJoCo. The demos use kinematic grasping. Neither these figures nor a control-frequency diagram establish physical safety.
The trillion-decisions/day and under-10-ms statements are five-year ambitions. Commercial tiers require a configured Stripe account; custom deployments require scoped onboarding.

Financial evaluation notes

The deck summarizes 11 wins or ties in 13 locked financial cells, five statistically significant wins, and no significant losses. This combines specialist versions and heterogeneous metrics; it is not a single composite score. The supplied visual and narrative do not provide per-cell confidence intervals for all 13 cells, so the site does not invent a forest plot.

The FOMC narrative reports 0.708 versus 0.532 and a +0.175 unrounded advantage. The displayed table computes +0.176 from rounded values. BigData lists Sonnet at +0.034 in the deck, while the copied raw run lists masked MCC +0.090; these baselines should not be treated as interchangeable.

The Q1 Herculean window reports about −1.7% versus −11.1%. The raw main-window daily-return difference has a confidence interval crossing zero. Post-window and pooled performance differ, so the selected period does not establish reliable trading alpha. C2B’s 300/300 exposure matches establish policy reproduction, not profitability.

Trading and market results are historical research simulations. They are not investment advice, a forecast, or a live trading service.

Brand resources

Kopyttse brand kit PDF · Icon SVG · Icon PNG · Favicon SVG · Favicon PNG · Full mascot SVG