Salvatore Sanfilippo (antirez), the creator of Redis, has published DwarfStar4 (DS4), a purpose-built C + Metal inference engine that runs the DeepSeek V4 Flash model locally on a 128 GB MacBook.
What it is and what it does
DeepSeek V4 Flash is a 284 billion-parameter open-source model with a 1,000,000-token context window. DwarfStar4 is not a general-purpose runtime; it is a dedicated engine optimized specifically for this model to extract performance beyond what generic tools provide.
According to the release, DS4 achieves about 27 tokens per second on a 128 GB MacBook.
Techniques and architecture
- C + Metal implementation: the engine avoids a generic runtime or framework and runs the model using raw C code and Apple Metal.
- Asymmetric 2-bit quantization: this compression method reduces the 284B-parameter model to roughly ~81 GB while reportedly preserving code generation and tool-calling quality.
- Disk-based key-value (KV) cache: DS4 includes an SSD-backed KV cache that stores session state. This helps agent workflows by avoiding repeated expensive prefill operations when clients resend large system prompts.
Integrations and ready APIs
DwarfStar4 ships with server APIs compatible with OpenAI and Anthropic interfaces, and provides ready-to-use configurations for agent environments such as Claude Code, opencode, and Pi. The release also mentions integration points for OpenClaw and Hermes.
Why this matters
The DS4 release is notable because it demonstrates that a highly opinionated, model-specific engine can run a quasi-frontier, very large model reasonably on a consumer-class but high-memory laptop. This approach may appeal to developers and researchers who need local execution, low latency, or resource-efficient agent operation without the constraints of general-purpose frameworks.
Key numbers and features
- Model: DeepSeek V4 Flash
- Parameters: 284 billion
- Context window: 1,000,000 tokens
- Target platform: 128 GB RAM MacBook
- Reported throughput: ~27 tokens/second
- Compression: asymmetric 2-bit quantization, model footprint ~81 GB
- Cache: disk-based KV cache on SSD
- API compatibility: OpenAI- and Anthropic-style server APIs
- Prebuilt configs: Claude Code, opencode, Pi; plus mentions of OpenClaw and Hermes integrations
The release presents a focused solution for running DeepSeek V4 Flash locally with a minimal, highly optimized engine designed for that single model, delivering usable performance on 128 GB MacBooks according to the published details.
