DGX Spark · documentation

Changelog

CHANGELOG.md Last pushed 3 September 2026

What a version means here: this repository is not a library, and nothing imports it. What you depend on is the defaults the recipe ships and the measurements taken on them. So a release is a measurement epoch — the configuration as it stood, and the figures that belong to it.

  • MAJOR — the measurement basis changes (new hardware, a different model or pack).
  • MINOR — a shipped default changes, or the recipe gains a capability. Your numbers move.
  • PATCH — documentation, corrections, tooling. Your numbers do not move.

Every entry leads with Defaults that changed. ./run.sh and the container print the version they were launched from. The image is tagged with it: ghcr.io/0xbakeer/deepseek-v4-flash-spark:<version>.

v0.1.1 — 2026-09-03

Defaults that changed

None that move a number. template_vars_default no longer pins thinking: false; the template derives the switch instead, and the rendered prompt is byte-identical for every request shape v0.1.0 accepted.

Changed

  • Thinking can be switched on with the OpenAI-style fields alone: a top-level reasoning_effort of medium/high/max (xhigh accepted as max), enable_thinking: true, or reasoning: {"effort": ...}. Before, only chat_template_kwargs.thinking did, and a client that sent reasoning_effort: high (Open WebUI's Reasoning Effort setting, for one) got the effort text with thinking still off. none/low keep the default chat mode; an explicit chat_template_kwargs.thinking still wins.
  • Docs updated accordingly (API, configuration, chat template, Open WebUI).

v0.1.0 — 2026-09-02

First release.

Defaults that changed

Everything — this is the baseline. The configuration as shipped:

Setting Value
pack anoane/DeepSeek-V4-Flash-0731-exl3-2.54bpw @ 91cf8b7e
engine exllamav3-anemone @ c6d2e3e + aarch64 patch
server TabbyAPI @ 4a4f9f4 + unified-memory patch
context 393,216 per request, 786,432-token cache pool, FP16
memory budget 106 GB main + 8 GB draft
draft MTP head, draft_mode: mtp
sampling temperature 1.0 / top_p 1.0 (DeepSeek's recommendation), overridable per request
reasoning off by default; chat_template_kwargs.thinking turns it on; effort low/high/max
tools tool_format: deepseek_v4

Measured on it

See bench/results.md and the table in the README. Summary: ~0.6 s to first token on a short prompt; ~30 tok/s streamed prose; 53–57 tok/s streamed code; prefill ~550–800 tok/s; 300k-token needle retrieved in a 374 s prefill.

Every number here was measured. Open an issue if one looks wrong.

All documentation