Skip to content

SparkGLM 0.3.6

Latest

Choose a tag to compare

@data-angel data-angel released this 18 Nov 18:22

Qualified and promoted 2026-11-18 on DGX Sparks; this is the release to install on two Sparks and on four. It fixes forced tool calls that could come back as text, and leaves GLM-5.3-Flash's output as 0.3.5 produces it.

Before promotion it was installed through LLooM's recipe at the recipe's own memory settings, with output identical to the qualification run; its Strict output and its image and video output were identical to 0.3.5's; and on a four-Spark chain its output was identical to 0.3.5's with speed level (results). The four-Spark run used GPU utilization 0.69 where the recipe sets 0.88, and Strict mode and images were not run on four Sparks.

SparkGLM 0.3.6 is 0.3.5's engine plus one fix:

  • A forced tool call opens with the native token (Enntity/atlas#33, upstream Atlas-Inf/atlas#395). With tool_choice set to required or to a named tool, the engine masks the first token of the answer so that it starts a tool call. The mask accepted the opening tag as the native <tool_call> token or as the same characters in plain text. A model that ranks the text form first then reasoned from that <, and the call came back as ordinary text with no tool call in the response. Only the native token is legal there now.
  • GLM-5.3-Flash FP4-Local is unchanged. It already opened forced calls with the native token.
  • Nothing else changes: tool_choice: auto, response_format with tools, and every token after the first are as before.

The fault showed on Sage S2.1 Flash, a checkpoint of the same architecture: 12 of its 109 live checks failed on 0.3.5, all forced-tool cases on the Anthropic API, and none fail on 0.3.6.

Against 0.3.5 on the same pair, Spark mode, profile 4x1m:

Check Result
Greedy output (8 prompts) identical
Prompt logprobs (20,705 tokens) zero difference
Live suite, GLM-5.3-Flash 109/109
Live suite, Sage S2.1 Flash NVFP4 109/109 (0.3.5: 97/109)
RigMark prose / code / structured (tok/s) 37.7 / 66.1 / 88.8 (0.3.5: 37.6 / 65.9 / 87.7)
Matrix walls c1-16k / c1-32k / c2-16k / c2-32k / c4-16k (s) 15.3 / 21.7 / 25.4 / 36.7 / 46.4 (0.3.5: 15.3 / 21.7 / 25.5 / 36.7 / 46.7)

Three things to know about these checks:

  • They ran at GPU memory utilization 0.90 under a 113 GB budget, where the recipe sets 0.91 and 114 GB. After the image build the head had about 0.3 GB less memory available than LLooM's admission asks for at 114 GB. The key-value pool was smaller by that much; none of these checks fills it. The install through the recipe then ran at 0.91 and 114 GB with identical output.
  • The speed and matrix rows are one run of the usual harness, not three.
  • Sage in Strict mode passes 97/109: three prompts that ask for tool markup repeated as plain text fail, each sent four ways. Serve Sage in Spark mode.

Receipts: results/2026-11-18-0.3.6.

Identity Value
Product commit 5e36d6bc23dbb862a4a1a415d9c697fc7f299de1
Install tree 9d6a1889555233cd8359db29230b566a645d1c2a
Atlas commit 01d34b343be6d675d8a1d6185e5b044b804b900c (Enntity/atlas spark)
Atlas tree f769ec3c1b741f490d1a561e4793b622388586b8
Image ghcr.io/enntity/atlas-sparkglm@sha256:e241d69dc3d1c592c6c605ad8e01b67763ff96178676282d40ccea18f43373d8 (tags 9d6a18895552, 0.3.6)

The image was built with the image checks off, as earlier releases were.

Install through LLooM (SparkGLM recipe version 7, the current version since Enntity/lloom#143):

lloom setup --recipe linux-nvidia-dgx-spark-glm53-atlas --additive --apply --yes