Strategic roadmap · Updated September 2026
2027 Roadmap
Rootcomputer’s plan for 2027, structured around early infrastructure expansion, consolidation of the model line, and novel architectural implementations reaching production — now with two milestones running ahead of schedule.
Progress
Where we are against the plan
H3 and our first CoFRNet model, M1, have both moved from the final quarter of 2027 into the first. Three legacy models retire earlier than planned to free hardware for that work.
- 2 Ahead of schedule
- 1 Date set
- 4 On track
- 3 Retired earlier
- Original plan
Model releases
Epos 3 Oct 1
Launch date set — October 1, 2026
Tanka T4 On track
January 2027
H3 Ahead
M1 · CoFRNet Ahead
Epos 3.1 & 3.2 On track
Q2–Q3 2027
Koda 2 On track
Q4 2027
Retirements
Epos 1.3 Earlier
Koda 1 Earlier
H2 Earlier
Epos 2.1 On track
Year-end 2027
Phases
The plan in four phases
Each phase pairs the model changes shipping in that window with the infrastructure work that has to land first.
Q4 2026 Added in this update
Epos 3 launch & hardware reallocation
Model releases & deprecations
- October 1 — Launch Epos 3.
- October 1 — Sunset Epos 1.3, Koda 1, and H2. Moved earlier
Infrastructure & technical R&D
- Reallocate freed hardware to next-generation pretraining and RL.
- Complete the H3 pretraining run.
- Continue M1 development toward Q1 beta trials.
Q1 2027
Compute acquisition, H3 & M1 beta
Model releases & deprecations
- January — Launch Tanka T4, renamed from Tanka 3.5.
- Mid–late Q1 — Release H3, succeeding H2. Ahead of schedule
- Mid–late Q1 — Begin M1 beta trials. Ahead of schedule
Infrastructure & technical R&D
- Secure and configure expanded training compute.
- Update the public site, documentation, and literature for Tanka T4.
Q2 – Q3 2027
Scaled training & flagship tuning
Model releases & deprecations
- Flagship — Maintain Epos 3, shipping minor versions via DPO and retraining.
- In development — Koda 2 training runs.
Infrastructure & technical R&D
- Scale parameter training across the new compute fleet.
- Validate CoFRNet FFN and PadéNet integration.
Q4 2027
Lineup refresh & legacy sunset
Model releases & deprecations
- Launch — Koda 2, succeeding Koda 1.
- Year-end — Sunset Epos 2.1 as extended support ends.
Infrastructure & technical R&D
- Production rollout of custom FFN architectures.
- Migrate public API and web traffic to the next-generation lineup.
Model lifecycle
Model lifecycle management
Clear support boundaries retire legacy builds while the flagship line stays continuous.
- Flagship continuity. Epos 3 launches on October 1, 2026 and remains the central flagship model through 2027, backed by iterative DPO post-training and minor version retraining runs (Epos 3.1, 3.2).
- Entry restructuring. H3 is a scaled H2: the same architecture at roughly 416M parameters, pretrained on 14.5B tokens against H2’s 4B. It replaces H2 at the base of the lineup in Q1 2027.
- Mid-tier restructuring. Koda 2 increases capacity over Koda 1.
- Sunset schedule. Epos 1.3, Koda 1, and H2 retire together on October 1, 2026, freeing hardware for continued next-generation training and RL. Epos 2.1 follows at year’s end 2027.
- Epos 1.3
End of Q1 2027Sunset October 1, 2026- Koda 1
Q4 2027Sunset October 1, 2026 — succeeded by Koda 2 in Q4 2027- H2
Q4 2027Sunset October 1, 2026 — succeeded by H3 in Q1 2027- Epos 2.1
- Extended support ends — year-end 2027
Architecture & compute
Architectural R&D and compute
Compute is acquired early so that scaling work mid-year is not bottlenecked by hardware.
- Early capital allocation. Concentrating compute acquisition in Q1 prevents training bottlenecks during mid-year scaling.
- M1 architecture. M1, the in-development name for our first CoFRNet model, replaces standard 2-layer FFN blocks with a modified CoFRNet FFN (continued fraction network) coupled with proprietary PadéNet rational approximant layers, aiming for higher expressiveness per parameter.
- M1 corpus. M1 is pretrained on 16.2B tokens that have passed through deduplication and denoising pipelines developed in-house, prioritising token quality over raw volume.
| Model | Architecture | Parameters | Pretraining tokens | Status |
|---|---|---|---|---|
| H2 | Transformer | 217M | 4B | Sunset Oct 1, 2026 |
| H3 | Transformer (scaled H2) | ~416M | 14.5B | Release mid–late Q1 2027 |
| M1 | CoFRNet FFN | 502M | 16.2B | Beta mid–late Q1 2027 |
This roadmap describes current intent, not committed delivery dates. Model names, versions, and timelines may change as research and training runs progress.