# DwarfStar (ds4): community site

> Community documentation, benchmarks and news for ds4
> (DwarfStar 4), Salvatore Sanfilippo's (antirez) local inference engine,
> supporting DeepSeek V4/V4.1, GLM 5.x and Qwen3.8 Flash Next on local
> hardware. Written in C with Metal, CUDA and ROCm backends.
> MIT license. Canonical implementation repository:
> https://github.com/antirez/ds4

Key facts:
- Supported models include DeepSeek V4 Flash (0731), DeepSeek V4 PRO (0813),
  DeepSeek V4.1 Flash, GLM 5.x and Qwen3.8 Flash Next, using only the GGUF
  layouts listed by upstream.
  Support is intentionally narrow and opportunistic rather than generic.
- Runs Qwen3.8 Q2 on a 64GB Mac: 41.73GiB of main/MTP weights are resident
  while the 95.37GiB BF16 n-gram table stays on fast local SSD.
- DeepSeek V4 Flash remains the 96GB+ baseline and can also run on 64GB Macs
  via Metal SSD streaming with reduced generation speed.
- DeepSeek V4.1 Q2 streams on one 128GB Mac, DGX Spark or Strix Halo (ROCm,
  text only), or runs resident across two Macs/two Sparks. Since 19 October
  2026 the ds41f-q2 target fetches DeepSeek's 1017 checkpoint
  (DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf); ds41f-mxfp4 is a 1017
  MXFP4 build (467GiB file). Q4 is the previous release and needs streaming or
  a 512GB Mac.
- GLM 5.2 runs on Metal, CUDA and ROCm, with mode-specific restrictions.
- GLM 5.3 Flash, Qwen3.8, DeepSeek V4.1 and DeepSeek Vision Experimental have
  matching vision paths. Backend support differs; V4.1 vision is Metal-only.
- DSpark speculative decoding (DeepSeek's draft model) is an opt-in runtime
  for Flash 0731 (resident only, including two DGX Sparks over network TP)
  and, with a separate 8.04GiB user-converted drafter, for DeepSeek V4.1 on
  Metal, CUDA and ROCm. GLM and Qwen use built-in MTP blocks instead.
- Pipeline parallelism splits layers across machines; two-Mac and CUDA tensor
  parallelism split model work differently and should not be conflated.
- ds4-server can keep multiple resident KV sessions and batch model-specific
  Metal or CUDA execution paths, with ordered fallbacks elsewhere. Qwen on
  Metal can also batch MTP speculation.
- Asymmetric 2-bit quantization: routed experts compressed, critical path
  kept high precision.
- KV cache persists to disk, keyed by SHA1 of the rendered prompt prefix;
  sessions survive restarts.
- ds4-server exposes OpenAI-compatible (/v1/chat/completions, /v1/responses)
  and Anthropic-compatible (/v1/messages) endpoints; works with OpenCode,
  Claude Code, Codex CLI and Pi.
- Native coding agent (ds4-agent) with persistent sessions; the project is beta.

## Docs
- [Quickstart](/docs/quickstart/): clone, download weights, build, run
- [Installation](/docs/installation/): hardware matrix, model variants, build targets
- [The CLI](/docs/cli/): chat commands, thinking modes
- [DeepSeek V4 Flash and PRO](/docs/deepseek-v4-flash/): builds, where it runs, DSpark, vision, PRO
- [DeepSeek V4.1 Flash](/docs/deepseek-v4-1-flash/): streaming, TP, thinking levels, vision
- [Qwen3.8 Flash Next](/docs/qwen-3-8-flash-next/): 64GB setup, MTP, vision, batching
- [GLM 5.3 Flash & vision](/docs/glm-5-3-flash/): artifacts, MTP, vision encoder
- [Server & API](/docs/server-api/): endpoints and client configs
- [The agent](/docs/agent/): session commands, KV persistence
- [Advanced](/docs/advanced/): SSD streaming, pipeline/tensor parallelism, batching, DSpark
- [Architecture notes](/docs/architecture/): quantization, KV design, validation
- [FAQ](/docs/faq/)

## Data
- [Benchmarks](/benchmarks/): ds4-bench context-frontier tables from upstream docs/PERFORMANCE.md
- [Blog](/blog/) and [RSS](/rss.xml)
- [GLM 5.2 is now a ds4 target](/blog/glm-5-2-and-the-next-target/):
  supported layouts, validation and execution-mode boundaries
- [GLM 5.3 Flash lands in ds4](/blog/glm-5-3-flash-lands-in-ds4/):
  the five-day support sprint and the project's first vision path

Benchmark data follows the upstream performance guide. Implementation details, model
targets, APIs and hardware guidance live in the ds4 repository.