# How DeepSeek Runs a 284B LLM on a Laptop (Run AI Locally)

> DeepSeek's 284B MoE model runs locally at ~26 tokens/sec on a 128 GB MacBook via a pure-C engine and asymmetric quantization: 2-bit experts, 8-bit router.

Published: 2026-06-25
URL: https://daniliants.com/insights/how-deepseek-runs-a-284b-llm-on-a-laptop-run-ai-locally/
Tags: deepseek, local-llm, mixture-of-experts, quantization

---

## Summary

DeepSeek's V4 Flash (284B params, MoE, MIT license) now runs fully local at ~26 tokens/sec on a 128 GB MacBook Pro M3 Max via Antirez's pure-C inference engine DS4 ("Dwarf Star 4"). The trick is asymmetric quantization plus mixture-of-experts: only ~13B params fire per token, and the engine crushes redundant expert layers to 2-bit while keeping the routing layer at full 8-bit. This is the first time a frontier-scale open model became a credible default for flash-tier work on hardware you own, with zero per-token API cost.

## Key Insight

**The architecture is what makes it possible**

- 284B total params, but MoE routing fires only ~13B per token, giving the knowledge of a huge model at the compute cost of a small one.
- Full-precision 284B model wants ~500 GB RAM. No laptop has it, so you quantize.

**Asymmetric quantization is the actual breakthrough**

- Naive approach: quantize everything equally, accept the quality hit.
- Antirez went asymmetric: expert layers (the bulk of params) crushed to **2-bit**; routing layer, shared expert, projection layer kept at **8-bit**.
- Logic: experts are redundant (many of them, few used per token) so fidelity loss is survivable. The router has no backup, so a bad route makes every downstream answer wrong. Spend bits where a mistake is fatal.
- Result: fits in ~76 GB, clocked at **26.68 tokens/sec** (independent review), faster than you read, on battery.

**Correctness-over-flexibility design choice**

- DS4 will NOT load arbitrary GGUF files; it ships its own validated weights tested against DeepSeek's official outputs. Critics call it lock-in.
- Antirez's counter: "run anything" is exactly why local inference stayed a hobby, since quality is undefined and you're guessing. Owning the whole pipeline (one model, one quant, one engine) lets him guarantee it works as claimed. Toy vs production.
- Not a thin wrapper: HTTP API, tool calling, built-in coding agent, disk-persisted KV cache (conversation memory survives across sessions).

**The honest asterisks (framing oversells)**

- Hardware floor is real: 128 GB MacBook Pro M3 Max = ~$3-4k; DGX Spark is a premium box. Democratization has a price floor well above most developer desks.
- "Quasi-frontier" is doing work: 2-bit experts cost quality; the claim comes from impressions, not rigorous head-to-head evals. Reviewers flag KV-cache edge cases and a silent quant-degradation trap.
- This is V4 **Flash**. The full V4 Pro is **1.6 trillion** params and needs very-high-memory machines, not a laptop. Frontier still lives in the data center.
- DeepSeek is a Chinese lab, so open weights solve privacy (data never leaves the box) but not training-data provenance or supply-chain/geopolitics questions.

**Why it mattered to people who'd know:** Georgi Gerganov (llama.cpp), the YC CEO, plus Metal/CUDA/ROCm backend ports appearing within weeks. Antirez's own tell: the first time he reached for a local model for serious work he'd normally hand to a frontier cloud model.