AI tool comparison
LazyMoE vs Ternary Bonsai
Which one should you ship with? Here is the side-by-side panel verdict, pricing read, reviewer split, and community vote comparison.
AI/ML Models
LazyMoE
Run 120B MoE models on 8GB RAM, no GPU, using lazy expert loading
50%
Panel ship
—
Community
Free
Entry
LazyMoE is an open-source inference engine built by a master's student in Germany that claims to run 120-billion parameter Mixture-of-Experts LLMs on 8GB of RAM with no GPU — using a technique called lazy expert loading. Instead of loading all MoE experts into memory at startup, LazyMoE identifies which experts are needed for each token at runtime and loads only those from SSD storage, keeping memory usage proportional to active expert count rather than total model size. The system is combined with TurboQuant KV compression (reducing KV cache memory footprint) and SSD streaming to minimize I/O latency when swapping experts. The builder demonstrated the system running on an Intel UHD 620 integrated graphics laptop — the kind of hardware that would typically struggle with a 7B model, let alone 120B. Token generation speeds are slow (a few tokens per second in the demo), but functional. If the claims hold up to independent testing, LazyMoE represents a meaningful democratization milestone: frontier-scale MoE inference made accessible on consumer hardware that most working professionals already own. The project is early-stage and from an individual researcher, so independent benchmarking is essential before drawing conclusions.
Open Source Models
Ternary Bonsai
1.58-bit LLMs that run at 82 tok/s on M4 Pro and on your iPhone
75%
Panel ship
—
Community
Free
Entry
PrismML's Ternary Bonsai is a family of aggressively quantized language models that take the BitNet concept to its logical extreme. Each weight is constrained to one of three values — {-1, 0, +1} — with a shared FP16 scale factor per 128-weight group. No higher-precision escape hatches, no hybrid layers. The result is a 9x reduction in memory footprint versus standard 16-bit models. The numbers are striking: the 8B model fits in 1.75 GB and hits 82 tokens per second on an M4 Pro. More impressively, it runs at 27 tokens per second on an iPhone 17 Pro Max — fast enough for real-time conversation on-device. The 8B variant scores 75.5 average across standard benchmarks, outperforming many models that are 9-10x larger. The 4B and 1.7B variants push further into mobile-optimized territory. All three models are released under the Apache 2.0 license, available on Hugging Face and GitHub, and integrated into the Locally AI iOS app for immediate on-device deployment. For developers building privacy-sensitive applications or anyone tired of paying cloud inference costs, Ternary Bonsai offers a compelling on-device alternative that doesn't require a beefy GPU.
Reviewer scorecard
“The lazy expert loading insight is genuinely clever — MoE models are already sparse by design (only 8-16 experts active per token), so you're not actually cheating, you're just not pre-loading experts you provably won't use. If the SSD throughput holds up on real workloads, this is the most practical approach to consumer-hardware frontier inference I've seen.”
“82 tokens per second on M4 Pro in 1.75 GB is a genuinely impressive engineering achievement. For local tooling, code assistants, or any latency-sensitive workload where I don't want cloud round-trips, this hits a sweet spot that larger quantized models miss. Apache 2.0 means I can embed it in commercial apps without legal headaches.”
“The demo shows a few tokens per second on a laptop — that's about 10-20x slower than usable inference speeds for most workflows. SSD read latency is also highly variable depending on hardware, and NVMe vs SATA would produce very different results. This is an interesting research demo, not a production inference engine. Also: master's student projects on GitHub deserve healthy skepticism about benchmark validity.”
“A 75.5 benchmark average sounds good until you compare it against 8B models quantized with GGUF Q8 — which score similarly and have years of tooling, community support, and production deployments behind them. The 9x memory savings matter on constrained devices but less so on any machine with 16GB+ RAM. Niche but real use case.”
“The trajectory here is clear: frontier-scale inference will become accessible to commodity hardware within 2-3 years, and techniques like lazy expert loading are part of how we get there. Even if LazyMoE itself is rough, the underlying approach will show up in production frameworks. This is worth watching as a proof of concept.”
“On-device AI at 27 tokens per second on a phone is the inflection point that makes LLMs a platform primitive rather than a cloud service. Once inference is this cheap and fast on commodity hardware, the entire economic model of AI-as-API-call collapses. Ternary quantization is an early signal of where efficiency research is heading.”
“Until token generation speeds reach at least 20-30 tokens per second, this isn't practical for creative workflows — writing, image generation assistance, or real-time collaboration. The technology is fascinating but the current demo is a proof of concept, not a working creative tool. Check back in six months.”
“The prospect of running a capable LLM entirely on my iPhone without sending any data to a server is genuinely exciting for creative work with sensitive material. Drafting, editing, and ideation without a cloud subscription or privacy concerns — I'd pay for that, and here it's free.”
Weekly AI Tool Verdicts
Get the next comparison in your inbox
New AI tools ship daily. We compare them before you waste an afternoon.