AN1: Meaning-First Compute for Hyperscale Inference
We compress transformer fields by 16× to 224× with near-zero accuracy loss, then run those compressed models 6× to 30× faster with a CUDA path and managed cloud service.
- •16× to 224× field compression at near-parity accuracy
- •6× to 10× faster inference out of the box, 15× to 30× in optimized stacks
- •Open source core, CUDA turbo path, fully managed cloud
What AN1 Actually Does
Reads fields from frozen transformers
AN1 taps intermediate hidden states from multiple layers of a frozen transformer and stitches them into a multi-anchor field.
Compresses them to a 256-dimensional meaning vector
It learns a 256d representation that preserves task-relevant structure, shrinking fields by 16× to 40× and up to 224× on frontier models.
Runs your workloads on the compressed field
A simple head on top of the compressed field nearly matches or slightly beats the full field classifier, while cutting compute and memory massively.
Field Compression Invariance
This is an empirical pattern we are observing: transformer intermediate layers contain far more information than their output dimensions suggest, and that information can be compressed dramatically while preserving task performance. We are not claiming this as a final law of nature, but the consistency across benchmarks is striking.
Proof that Compression Works
| Task | Dataset | Compression | Baseline | AN1-256d | Δ |
|---|---|---|---|---|---|
| Sentiment | SST-2 | 40× | 90.9% | 91.4% | +0.5% |
| Entailment | MNLI | 40× | 70.9% | 71.3% | +0.4% |
| Yes/No QA | BoolQ | 40× | 68.3% | 68.4% | +0.1% |
| Commonsense | HellaSwag | 16× | 83.08% | 83.36% | +0.28% |
| Llama-3.3-70B | SST-2 | 224× | 92.78% | 93.35% | +0.57% |
Interpretation
AN1 achieves lossless semantic compression at 224× while slightly outperforming the raw 70B teacher activations. This indicates that Llama-70B's semantic structure resides on a compact, low-rank manifold that AN1 captures more efficiently than the uncompressed representation.
AN1 Stack
AN1-Core
Open for validation, closed for production.
AN1-Core began as an open research layer to enable independent validation and peer review. The production AN1 Engine is now closed-source and commercially licensed.
AN1-Turbo
Proprietary CUDA kernel stack
- •Fused kernels for field projection
- •Optimized 256d matmuls
- •Quantization and batched inference
- •6–10× out of box, 15–30× fully optimized
Anima Cloud
SoonManaged service layer
- •Fully hosted AN1-Turbo
- •Usage-based pricing tied to compute savings
- •Green credit participation
Patent pending. AN1 is powered by proprietary semantic-field technology developed by Anima Core Inc. US provisional patents have been filed to protect the core innovations.
Pricing model: performance aligned
AN1 licensing is structured around shared value. Clients pay a percentage of the compute they save, ensuring every deployment is a net positive for cost and energy.
Cloud providers
5–12% of verified savings
10–20% participation in green credits from reduced GPU energy
Hyperscaler savings reach eight to nine figures annually
Large enterprises
8–15% of compute savings
Optional green credit co-registration
Replaces traditional mid-six to low-seven figure licenses
Smaller enterprises
10–18% of compute savings
Flexible monthly or quarterly billing
For teams spending five to six figures per year on GPU workloads
Nobody pays until they save. The model aligns incentives: the more performance AN1 delivers, the more both parties benefit.
Anima Cloud
Launching SoonFully managed AN1-Turbo inference as a hosted service.
For teams that prefer not to operate GPU infrastructure, Anima Cloud delivers AN1-Turbo acceleration through a fully managed API with the same 6–10× speedup out of the box and 15–30× in optimized stacks.
What you get
- •Zero infrastructure overhead
- •Simple REST or gRPC API
- •Green credit participation
Pricing
Pay 8–15% of verified compute savings. If you would have spent $10,000/month on uncompressed inference and AN1 reduces that to $2,000, you pay ~$640–$1,200 of the $8,000 saved.
Honest Roadmap
Q1 2026 — Pilot Launch
- • First AN1 pilot deployments with select enterprise and cloud partners
- • Integration of AN1-Core + early AN1-Turbo acceleration path
- • Real workload benchmarking and cost validation
- • Feedback loop for CUDA kernel refinement
Q2 2026 — Turbo Acceleration + Kernel Deepening
- • Advanced CUDA kernels for field projection and 256d inference
- • End-to-end Turbo acceleration benchmarks across multiple model sizes
- • Expanded pilot program with additional partners
- • Joint validation reports with infra teams
Q3 2026 — Anima Cloud Private Beta
- • Launch of Anima Cloud private beta
- • Usage-based savings meter, dashboards, and observability
- • Automated deployment pipeline for AN1 inference
- • Green credit integration pilot with select partners
Q4 2026 and Beyond — General Availability
- • Public release of AN1-Turbo and Anima Cloud managed service
- • Enterprise-grade SLAs and support
- • Large-scale deployments and commercial agreements
- • Continued acceleration of symbolic compute research
Get in Touch
Interested in AN1-Turbo licensing, Anima Cloud access, or partnership opportunities? We are actively working with hyperscalers, cloud providers, and large enterprises.