octa. An agentic foundation model for marketing, GTM, and CRM.
Built on top of a strong open-source base, fully developed in-house. A harness and an LLM working in tandem - both are critical. The training corpus is a decade of real campaigns from 1,000+ B2B brands across the CIENCE and Tenbound archives. From here, every customer campaign sharpens the next training pass.
A decade of real campaigns, structured for training.
Foundation models are bottlenecked by their training data. The public web teaches them prose. octa will be trained on real GTM outcomes. That is the moat.
Continuous campaign telemetry from 2015 onward. Multi-region, multi-industry, multi-channel.
Real B2B accounts across SaaS, finance, healthcare, manufacturing, media, public sector.
Email, voice, SMS, social, chat, landing pages. Every variant logged with the outcome it earned.
Reps tagged the why behind every conversion and disqualification. octa will learn from the tags.
Why generic LLMs miss GTM.
GPT-5, Claude, Gemini are excellent at prose. They are not built for GTM. They have never watched a rep handle a real objection, never seen which 3 follow-up lines turn a reply into a meeting, never learned the cadence that beats the alternative by 11% in a specific segment.
octa is being built the other way around. The model starts from real campaign outcomes - emails, calls, replies, meetings, landing pages, deal closes - then learns to reason and write on top of that. Every output is grounded in a touch that historically did or did not work.
- Generic LLM: trained on the public web. Knows what a cold email looks like.
- octa: trained on real campaigns. Will know which cold email got the meeting.
From a thousand SDRs to one model.
A decade of CIENCE campaigns and years of Tenbound research generated the raw signal. Capchase contributed both archives into graph8 in 2025. graph8 is turning the combined corpus into octa.
- 2015
CIENCE founded
Outbound campaign service launches. Day one: log every touch, every reply, every meeting set, every disqualification reason.
- 2015 - 2024
A thousand SDRs, 1,000+ brands
Daily CIENCE campaigns across SaaS, finance, healthcare, manufacturing, public sector. Reps tag the why behind every win and loss. A decade of labeled GTM outcomes accumulates.
- 2025
graph8 platform launches
AI sales platform ships with the harness, the agents, the integrations. octa is on the roadmap.
- 2025
David Dulany joins graph8
David Dulany joins graph8 with his wealth of GTM research, sales development frameworks, and category expertise built at Tenbound.
- 2025
Capchase becomes a graph8 shareholder
Capchase joins as a shareholder of graph8 and contributes both the CIENCE and Tenbound campaign archives into the company. The two complementary GTM data lineages now sit inside graph8, ready to train octa.
- 2026
octa training begins
Pretraining starts on top of a strong open-source base, using the combined CIENCE + Tenbound corpus. First model heads target copy, qualify, pages. octa Bench is built from real anonymized campaigns.
- Underway
octa-mini, octa, octa-reasoning
Three model sizes in development: low-latency drafting, balanced agentic execution, deep reasoning. Early signal shows measurable lift over the base LLM. Public benchmarks coming when they survive evaluation.
The brands whose campaigns will train octa.
Over 1,000+ B2B brands ran campaigns through CIENCE between 2015 and 2024. The outcomes, the replies, the objections, the meetings booked, the landing pages that converted - all of it becomes the training corpus for octa. Top 5 per tier shown here for context. See the full case-study library at cience.com ->
Enterprise
1,000+ employees
Mid-Market
200 to 999 employees
SMB
50 to 199 employees
Startups
Under 50 employees
And 1,000+ more brands across the CIENCE archive. Browse the full case-study library ->
Ingest. Label. Continued pre-training. Long-horizon RL. Evaluate + ship.
A continuous loop. Real campaign events come in. Human reps tag the why. octa retrains. octa Bench scores. New weights ship to every customer org. The next campaign feeds the next loop.
Ingest
Every campaign event flows in: sends, opens, replies, calls, meetings, deal stages, landing page sessions, form submits. Multi-channel, structured, time-stamped.
Label
Human SDRs and AEs tag the why: which lines worked, which objections killed the deal, which landing page variant turned a click into a meeting.
Continued pre-training
Pick a strong open-source base model. Run a large-scale continued pre-training pass on the GTM corpus. Goal: deep knowledge of the specific work, not generic chatbot fluency.
Long-horizon RL
Rollouts of realistic GTM tasks: write the next message, plan the sequence, generate the landing page, qualify the reply. Length penalty and self-summarization keep long sessions on track.
Evaluate + ship
octa Bench replays held-out historical campaigns end-to-end. New weights ship to every graph8 org. Outcomes feed back into the corpus. The loop closes.
Every campaign sharpens the next one.
The day a customer turns on graph8, their campaigns start contributing to octa's training. Objections raised this week become labels for next week's fine-tune. Every reply, every meeting, every landing page conversion is a vote that updates the model.
Static prompts decay. octa compounds. A model trained on real GTM outcomes today is sharper than the same model a quarter ago, and dramatically sharper than a generic LLM that has never run a single campaign.
Three model sizes, one foundation.
Same training corpus, three deployment shapes. We pick the right model for the task. Latency matters in the inbox. Reasoning matters in the deal review. Both matter in landing page generation.
octa-mini
Low-latency drafting headSmall, fast, always-on. Drafts emails, fills sequence variables, scores reply intent in milliseconds inside the inbox and the dialer.
octa
Balanced agentic executionThe default workhorse. Plans sequences, qualifies replies, generates landing pages, writes follow-ups, books meetings.
octa-reasoning
Deep multi-step reasoningFor complex GTM work: account-level strategy, multi-channel orchestration, edge-case objection handling, deal-stage forecasting.
How we are building octa.
Five short slides on the work in progress. We are following a research arc that mirrors the broader field: pick a strong open-source base, run continued pre-training on the domain, use long-horizon reinforcement learning to shape behavior, build an internal benchmark from realistic queries, ship. This is what we are doing for GTM. None of it is finished. The slides describe the plan.
Deep GTM knowledge. Hard tasks to completion. Realistic motions.
octa is being built around three goals. Real domain knowledge of marketing, sales, and CRM work - not generic chatbot fluency. The ability to run hard end-to-end GTM tasks all the way to a booked meeting or a generated landing page. And winning on realistic motions our customers actually run every day, not contaminated public benchmarks.
Continued pre-training on a strong open-source base.
We start with a strong open-weight base model. We run a large-scale continued pre-training phase on the CIENCE corpus and the live graph8 telemetry. The point is not to add general chatbot polish - the point is to push the model toward the specific work it will be asked to do every day inside a GTM team. Short-context first, long-context extension second, supervised fine-tuning to round out the agentic shape.
Rollouts of real campaigns. Length penalty. Self-summarization.
The primary training phase will be long-horizon reinforcement learning. We collect a large set of realistic GTM problems: write the next touch, recover the stalled reply, plan the cross-channel cadence, generate the landing page, qualify the inbound. We run many rollouts per problem, score them on real outcomes, and update toward the winners. A nonlinear length penalty keeps short tasks short and lets hard tasks run long. Self-summarization lets the agent push past the context window for tasks that need it.
octa Bench - real anonymized campaign queries, not public leaderboards.
Public LLM benchmarks measure trivia and code puzzles. None of them measure whether the next message gets the meeting. octa Bench is being built from anonymized real queries pulled from the corpus - ill-structured the way real GTM requests are, with realistic ambiguity the agent has to resolve. We score against the outcome the human actually earned. The bench will separate models that are good at GTM from models that are only good at sounding plausible.
octa-mini today, octa next, octa-reasoning soon, more channels after.
The first three model sizes target the most common GTM shapes: fast drafting (octa-mini), balanced agentic execution (octa), deep reasoning (octa-reasoning). The next pre-training pass will scale the corpus to include voice transcripts at scale, full landing page DOM trees, and live ad creative. Goals come before benchmarks. We will publish numbers when they survive evaluation.
The work, the data, and the model are graph8's own. We will not be quoting numbers we have not earned.
What we are optimizing octa for.
We are not going to publish benchmarks we have not earned. Internal evaluations today show octa-mini, octa, and octa-reasoning each delivering measurable lift over the base open-source LLM on GTM tasks. The numbers are small but consistent and trending up as the corpus and labels grow. Full results when octa Bench is production-ready.
Numbers stay private until they survive a held-out evaluation against the corpus. Talk to our team for the segment that matches your motion - we will replay one of your past campaigns and show you what octa would have shipped.
octa vs general LLMs, by GTM task.
Same input, different upbringing. Generic LLMs read the public web. octa will watch a decade of campaigns and learn how to ship pages, not just paragraphs.
Picks the opener that this persona actually replied to in historical campaigns.
Writes plausible prose with no link to a real reply rate.
Recognizes the objection pattern, drafts the response that reopened the conversation in past data.
Apologizes generically. Loses the thread.
Knows the cadence that beats the alternative in this segment, and which channel to switch to mid-sequence.
Defaults to a generic 5-email drip. Same as every other vendor.
Generates the layout, headline, form, and proof block grounded in what converted for similar offers.
Generates copy. You still hire a designer and a builder.
Six specialized heads, one model.
octa is not a single chat model wrapped in a prompt. Each GTM job has its own head, trained on the slice of the corpus that matters to that job, then reinforced by its own outcome signal. Landing pages are first-class - generated as full pages, not just copy.
Research
Builds account briefs, persona docs, and competitor maps from the open web plus your private context.
Copy
Drafts emails, ads, sequences in your brand voice. Scored against historical reply rates from the corpus.
Pages
Generates full landing pages: layout, headline, form, proof. Not just text. Trained on what converted.
Qualify
Reads inbound replies and meeting transcripts. Scores fit. Routes to the right rep with context attached.
Coach
Reviews call recordings and reply threads. Surfaces what worked. Gives reps the next-best action.
Forecast
Predicts pipeline conversion from stage signals. Flags slipping deals before the close date.
What octa is doing right now.
Live across every graph8 customer org. Every interaction adds to the corpus that trains the next version.
octa pipeline is drafting, generating, and qualifying inside every graph8 customer org
Each specialized on one GTM job. Each retrained as new outcomes land.
Logged, scored, and folded back into the next training cycle.
Bring your motion. octa already knows the patterns.
Talk to our team. We will replay one of your past campaigns through octa, show you the lines, the cadence, and the landing page it would have shipped, and the meetings the model would have set.
This is the model story. The process story lives at octa² - the infrastructure of six algorithms running around the model. The two pages are a pair.
The brands referenced above engaged CIENCE or Tenbound for campaign services between 2015 and 2024. Both campaign archives were contributed into graph8 in 2025 via Capchase, when Capchase came on as a shareholder. David Dulany joined graph8 the same year with his Tenbound-built expertise. Logos are used for historical training-corpus reference under fair-use or with permission. Reference is not an endorsement of graph8 as a current platform. 1,000+ more case studies at cience.com/case-studies.