After thinking about the very high margins that Micron recently reported (greater than Nvidia’s) I asked Grok 4.5:
What startups are pursuing technology that reduces memory requirements for ai models and inference?
Several startups are actively developing technologies to slash memory requirements for AI models and inference. These efforts target the core bottlenecks in large language models (LLMs) and multimodal systems—primarily through extreme quantization (reducing bit precision of weights), alternative architectures that avoid quadratic memory scaling (e.g., state-space models or hybrids), sparsity/partial activation (like Mixture-of-Experts or neuron-level sparsity), efficient runtimes/kernels, and hardware innovations like in-memory computing that minimize data movement.
This space is especially active for on-device/edge AI (phones, laptops, embedded systems) and cost/energy-efficient cloud inference, where fitting larger or more capable models into limited RAM/VRAM, reducing KV cache bloat in Transformers, or lowering bandwidth demands delivers major gains.
Here are some of the most prominent startups and their approaches (as of mid-2026):
Extreme Quantization & Model Compression
- PrismML (Caltech spinout, backed by Khosla Ventures): Focuses on extreme low-bit architectures—1-bit binary (“Bonsai”) and ternary (1.58-bit, weights in {-1, 0, +1}). They compressed Alibaba’s Qwen3.6 27B model from ~54 GB (FP16) to under 4 GB and ran the full dense 27B model on an iPhone 17 Pro. Claims include up to 14× smaller memory footprint, 8× faster inference, and significantly lower energy use, with competitive or better benchmark performance. They have open-sourced Bonsai models (including smaller 8B/4B/1.7B variants) under Apache 2.0 and are in discussions with Apple. This represents one of the most aggressive commercial pushes into 1-bit/ternary models for on-device deployment.
- Mobius Labs (Berlin): Developed Half-Quadratic Quantization (HQQ), a fast, calibration-light post-training quantization method that enables high-accuracy low-bit models (including aggressive 2-4 bit). They demonstrated quantizing Llama 70B to run on a single GPU instead of four without major accuracy loss, directly cutting memory and compute needs. Their work extends to FP4 optimizations and integrates with frameworks like vLLM.
Alternative Architectures for Inherent Memory Efficiency
- Liquid AI (MIT spinoff): Builds Liquid Foundation Models (LFM / LFM2 series)—hybrid architectures combining gated short convolutions with grouped-query attention (GQA) blocks, plus MoE variants. These deliver substantially lower memory footprints than Transformers (especially for long contexts, avoiding massive KV cache growth), faster prefill/decode (up to 2× on CPU in some cases), and strong on-device performance. Examples include tiny models (230M–350M params, often
- Cartesia: Specializes in state-space models (SSMs) and Mamba-derived/hybrid architectures. These offer near-linear (or better) scaling with sequence length and near-constant memory complexity during inference (fixed or minimal growth vs. quadratic KV cache in Transformers). This enables efficient long-context processing on smaller devices with lower energy use. They provide on-device libraries (e.g., “Edge”) and models like hybrid Mamba-2 variants optimized for throughput and fixed memory footprints.
On-Device/Edge Runtimes & Optimizations
- Cactus (Y Combinator-backed): Builds cross-platform, energy-efficient inference runtimes and kernels for mobile, wearables, and low-power devices. Supports aggressive quantization (down to 2-bit), delivers very low latency (sub-50 ms time-to-first-token), and keeps models small (e.g., hundreds of MB for small models). Emphasizes privacy (fully local) and broad model support (Llama, Qwen, Gemma, etc.).
- Embedl: Focuses on edge-optimized multimodal models and mixed-precision quantization (e.g., W4A16-Edge2 with selective FP16 layers + “FlashHead” optimizations). Enables fast multimodal reasoning on devices with under 8 GB RAM (e.g., Jetson Orin Nano/AGX), with significant speedups over baselines while recovering most accuracy.
Hardware Innovations Reducing Memory Pressure
- d-Matrix: Develops digital in-memory computing (DIMC / 3DIMC) chips and platforms (e.g., Corsair) that integrate compute directly with/near memory (initially SRAM, now advancing to 3D DRAM). This drastically cuts data movement—the dominant memory bandwidth and energy bottleneck in inference. Claims include major speedups (up to 10× in some workloads) and better efficiency vs. traditional GPU/HBM setups, enabling higher performance with more manageable overall memory system requirements.
- Hailo: Designs efficient edge AI inference chips (e.g., Hailo-10H series) optimized for low-power LLM and multimodal workloads. Their chips target 1–3B parameter models at very low power envelopes (~2–2.5 W), making deployment feasible on memory- and power-constrained edge devices where traditional hardware would struggle.
Other Notable Mentions
- Tiiny AI: Uses neuron-level sparse activation (“TurboSparse”) + heterogeneous engines for running very large models (up to 120B) fully on-device/offline in compact hardware.
- Broader ecosystem players (inference platforms like Together AI or Fireworks) heavily optimize serving (quantization, caching, KV cache management) but are more infrastructure-focused than core technology innovators in compression/architecture.
Key Trends and Context
Quantization (especially post-training or quantization-aware) remains the most immediate lever and is widely adopted. Architectural shifts (SSMs, hybrids, MoE with low active parameters, convolutions) address fundamental scaling issues like KV cache memory growth. Hardware approaches (in-memory compute, high on-chip SRAM, efficient NPUs) tackle bandwidth and data movement. Many solutions combine these (e.g., architecture + quantization + sparsity).
Progress is rapid, driven by on-device AI demands (Apple’s interest in PrismML is notable) and the economics of scaling inference. Many models are open-sourced or have open weights, accelerating adoption. Results vary by workload—extreme 1-bit models excel in certain scenarios but may need custom kernels; SSM/hybrid models shine for long context or linear scaling needs.
For the absolute latest benchmarks or specific model deployments, check company sites (e.g., PrismML’s Bonsai releases, Liquid AI’s LFM models on Hugging Face) or recent evaluations, as the field evolves quickly.
Interesting!