For: Chris · CTO · Hypernym
2026-05-11 · drafted from R18 panel synthesis · prepared as R19 seed
Two architectural priorities + one product wedge + one developer-stack play, all converging on the inference-core thesis. Centered on the compound-3×-MTP × 3×-Modulum opportunity you flagged. Modulum API is currently live and responding at 200 OK on Q4_K_M Gemma 4 31B with active 2-token speculative drafting (verified 2026-05-11).
Modulum was developed pre-MTP. MTP changes the decode flow: each pass produces N tokens instead of 1. Modulum's per-pass attention savings compress across N tokens — the 3× decode advantage erodes when MTP is active. The architectural fix lives at the attention-mask level, not the decode-stage level.
The compound-erosion problem is concrete: MTP reduces pass count by ~N×; Modulum reduces per-pass attention cost by ~3×. They operate on different bottlenecks. When fused naively, the per-token win from Modulum compresses across N MTP-predicted tokens.
Three architectural paths that compound (not compete) with each MTP head:
k (predicting token t+k) gets its own Modulum mask tuned for position-k retention bandwidth. Head 1 = proximate retention; head 8 = distal retention. Compound speedup recovers because each head is locally optimal for its prediction horizon.3× MTP × 3× Modulum-fused = 9× compound decode at long context. Achievable in principle if the fusion happens at the attention-mask level. The fusion is the technical risk; the math holds if Modulum masks remain optimal per-MTP-head.
None — not all but few — frontier labs have MTP enabled at production scale today. The compound moat (Modulum + MTP) is winnable as a partnered training run with a frontier-lab partner, not as an inference-time-only retrofit. Both routes are valid; the JV path is the larger upside and the scaffolding for category-defining results across base-model families.
OpenRouter sits between developers and frontier APIs taking a unified-billing cut. ~100k developers use them as the single base_url. They don't add value beyond routing/billing/fallbacks. Hypernym does.
Hypernym Router (HR): drop-in base_url replacement for OpenRouter. Same OpenAI-compatible interface, same routing/billing/fallback, BUT every call routes through Modulum-enhanced inference where applicable. Customers pay similar to OpenRouter, get +9pp retention at 128k for free underneath.
OpenRouter charges ~5% over wholesale. HR charges 0-3%, subsidized by Modulum's decode savings as margin. Compound MTP × Modulum amplifies this.
HR is genuinely faster on long-context calls vs vanilla OpenRouter. Retention is the speed multiplier — fewer retries because facts don't get lost.
Every call returns a Retention Receipt (R18 convergent product). OpenRouter has nothing equivalent. Procurement-defensible.
Direct developer + small-startup market. No procurement, no contract. pip install hypernym-router or just the base_url swap. Wide-market reach by definition — exactly what we need for fast adoption while institutional pathways mature on slower clocks.
Modulum-retained evidence plus an explicit proof/claim/obligation graph that the model updates and verifies during generation. The final answer includes a verifiable dependency trace. This is R19's flagship research stream — but the Day 1 customer is concrete.
API: https://reasoning.hypernym.ai/v1 — returns answer + dependency trace + claim graph state · SDK: pip install hypernym-reasoning adds proof/claim/obligation tracking to any agent loop · Pricing: $0.10-0.50 per traced query (premium over base Modulum) · Day-1 customer: any developer building an agent that needs to remember what it did 30 steps ago.
Reasoning-state architecture becomes a developer product before an institutional product. Wide-market reach (agent developers, tens of thousands) immediately; regulated-institution adoption follows once the trace is procurement-defensible.
Forge is already a multi-track research orchestrator, LaunchAgent fleet manager, cross-model dispatch + cost tracking, and memory infrastructure. Extend it to also be the Modulum customer-deployment toolkit.
Customer flow: deploy → benchmark → integrate in a single forge session, all open-source, all reproducible. Adoption pipeline runs through forge; customers verify the +9pp retention gain in their own benchmarks before committing. Direct top-of-funnel into Hypernym Cloud or self-hosted Modulum deployment.
Forge becomes the developer-relations tool for Modulum. Customers who run forge to deploy Modulum locally become the warm leads for Hypernym Router, Hypernym Reasoning, and (eventually) institutional deals.
From R18 R1 (Grok origin, panel-converged). M5 patterns are partly architectural (transfer-friendly) and partly training-corpus-specific (transfer-unfriendly). Distill M5 into a 50M parameter "attention compiler" that emits per-layer sparsity for any new model in minutes.
You're already running Llama 4 infrastructure per-head-scale. The transfer experiment is cheap — warm-start from Gemma 4 31B patterns; calibrate on a single BABILong qa1 sweep at 128k; measure retention gain. Predicted floor: +5pp on Llama 4 70B, +4-6pp on Mistral Large 3, +5-7pp on Qwen 3 72B.
If M5 patterns are mostly model-agnostic (attention entropy profiles rather than weight-specific), the moat scales horizontally. Hypernym ships compile.hypernym.ai as a developer service — any new frontier or OSS model launches; we have Modulum-on-it running within a day. Counter to a frontier lab's "we built our own retention layer" — ours composes with theirs.
Climate is hard, expensive, and slow to validate. Cheaper world-model targets exist where Hypernym's substrate-grounding moat applies directly and validation is tractable.
| Domain | Why cheap + tractable | Spend estimate |
|---|---|---|
| Economic / financial market simulators | Real-time data is free (yfinance, FRED, SEC EDGAR). Validation = backtest against historical prices. | $50K cloud compute / 6 months |
| Materials science (small-molecule + crystal) discovery | Open data (Materials Project, OQMD). Validation = predict synthesizable + DFT-stable. Pharma + battery + semiconductors TAM. | $100-150K |
| Logistics / supply-chain optimization | Synthetic + open data. Direct procurement appetite (every F500 has logistics teams). | $80-120K |
| Multi-agent game theory / mechanism design | Pure simulation, no real-world data costs. Crypto / DeFi applications. | $30-60K |
| Drug-protein binding simulators | AlphaFold-era data is open. Pharma customers pay seven figures. | $100-180K |
Economic-market simulators OR drug-protein binding as the first non-weather world-model bet. Both have clear validation (backtests / wet-lab experiments), both have direct enterprise customers, both fit the $200K cap. Both also let Hypernym productize substrate-grounded simulation distinct from climate's coupled-PDE complexity.
200 OK in ~1.2s smoke test 2026-05-11. Model: gemma-4-31B-it-Q4_K_M.gguf. Active 2-token speculative drafting with 100% acceptance on short prompts. Endpoint: https://gemma4.hypernym.ai/v1.
All canonical services up. Zero drift. Zero orphans. Research LaunchAgents on-demand. Memory router 10/12 providers healthy. Ready to deploy Modulum benchmarks + customer integrations.
R19 = "Modulum × MTP + Inference Stack." Expands R18's reasoning-state direction to include the MTP compound work + commercial wedges.
reasoning.hypernym.ai/v1; SDK adoption.forge deploy / benchmark / integrate; developer-relations pipeline.compile.hypernym.ai.Plus operational: R19 dispatch fixes (Gemini single-key auth path, Gemma mlx_vlm CLI invocation); IHC benchmarking integration as a sign-in page + proof doc URL + API access (no cost-tracking infrastructure — they'll do their own).
Don't waste time on cost-tracking infrastructure. They'll do their own.
Stand up a small webpage at benchmark.hypernym.ai (or similar) where IHC contacts:
Single-day buildout. Stripe + Clerk + the proof doc + the existing Modulum API. No telemetry, no cost-tracking — just hand them the keys.
The lost-in-the-middle results are the #1 thing Hypernym can hold its hat on right now. The proof doc validates +9pp at 128k against vanilla Gemma 4 31B on a public benchmark, with frontier labs visibly regressing 32-46pp on multi-needle at the same lengths. Hypernym is moving the right direction at 5 orders of magnitude lower cost.
R19's job: layer MTP-compound architecture + Hypernym Router commercial wedge + reasoning-state for agents + forge deployment toolkit + cross-model transfer + one cheap world-model bet on top of this proven foundation. All of these adopt by API call, none require institutional procurement.
Modulum stays the core thing. Hypernym Router shows the world how it works. Reasoning state extends it to multi-turn agents. Forge ships it to developers. Cross-model transfer makes it universal. The MTP compound work — most importantly — keeps the decode advantage growing rather than eroding as base-model architectures evolve.