Models

Open models, measured at every release.

Open language models, and the models that build and search a knowledge graph. Each is shown with the speed one user sees and the total speed across many users served at once, as the latest release measured them.

These figures are generated from the release record when the page is built. None is typed by hand, and a model that was not measured reads Results pending, never zero.

Early diagnostic speeds on the release gate's 2-node box (8 CPU cores per node), measured every release and being improved. They are not a benchmark against other engines.

Speeds in tokens per second, each from the newest published release that measured the model, named in its row; the latest published release is r11.127, 26 October 2026. One user: how fast a single conversation is answered. Many users together: the total across many conversations served at once.

Language and multimodal models

The models a person or an agent talks to. Only mixture-of-experts models are listed.

Model	Lab	Size	Licence	Status	One user	Many users together	Measured at	Accuracy	Resources
Gemma 4 26B-A4B	Google	26B total, 4B active	Apache 2.0	retired until the native cell path lands	not measured yet: its native serve is built and is being made the one production serve			Results pending	Results pending

Knowledge graph and retrieval models

The models that turn documents into a graph of facts and find what a question needs. As of 2026-10-10, Moonshot (Kimi), Thinking Machines and Mistral publish no open-weight embedding model on Hugging Face.

Model	Lab	Size	Licence	Status	One user	Accuracy	Resources
Qwen3-Embedding 0.6B	Qwen	0.6B	not stated	installable; not on the release gate's node	not measured yet: installs from the Marketplace; its speed is being measured on the current cluster	Results pending	Results pending
Qwen3-Embedding 8B	Qwen	8B	not stated	not on the release gate's node	not measured yet: its weights are not yet converted to Semurg's storage format, and this kind of model has no measure yet	Results pending	Results pending
Qwen3-Reranker	Qwen	not stated	not stated	installable; installed, not yet measured	not measured yet: installs from the Marketplace (off until enabled); its speed is being measured on the current cluster	Results pending	Results pending
Qwen3.5-9B	Qwen	9B	not stated	not on the release gate's node	not measured yet: its weights are not yet converted to Semurg's storage format, and this kind of model has no measure yet	Results pending	Results pending
EmbeddingGemma 2	Google	744M total, text tower 271M	Apache 2.0	installed by default; installed, not yet measured	not measured yet: installs by default from the signed Marketplace catalogue; its speed is being measured on the current cluster	Results pending	Results pending

Speeds: being re-measured

Every model on this list is being re-measured. Figures from before 13 October 2026, many from a test cluster since retired, are not shown. Until the release gate measures a model, its current speed is the one measured from outside on a fresh node of the published release, in the table below.

Measured from outside on the published release r11.86.20261013.1f4b26c (13 October 2026)

A fresh node of the published release, started from the run line on /develop, its models installed from the Marketplace and measured over HTTP from outside the node; every reply checked; three cycles (cold, warm, warm again), the last shown. Processor busy is the mean over all 16 hardware threads during the measurement. On one node: an 8-core / 16-thread AMD EPYC 4345P with 2-channel DDR5, every core available to the node.

Model	What was measured	Result	Processor busy
EmbeddingGemma 2	one passage per request, one client: latency p50 / p99	214 / 245 ms	34%
EmbeddingGemma 2	32 passages per request, one client	7.4 embeddings/s	59%
EmbeddingGemma 2	32 passages per request, 8 clients at once	10.5 embeddings/s	98%
Qwen3-Reranker	one query over 20 passages, one client: latency p50 / p99	7.7 / 7.8 s	66%
Qwen3-Reranker	one query over 20 passages each, 8 clients at once	0.18 queries/s	100%

Data and models, one store

Semurg is being built to serve a model from the same store as your data, on the same engine. What that is for:

Models larger than memory. A model's weights stay on disk and are read as the computation needs them, so a node can serve a model bigger than its RAM.
One system. The model and your memory live in one store, under one install, one licence and one set of access rules.
Faster with use. The engine learns which parts of a model are used together, so repeated work reads less.

Where this stands in the latest published release (r11.127, 26 October 2026): this is being built, and none of it is in a published release for models yet; each model's row above says what it is waiting for.

How to read these figures

Measured at a release. A figure appears only when a published release measured that model, and its row names that release. Nothing is estimated, and a figure is never carried from a build that was not published.
Not measured yet. A model that no published release has measured reads not measured yet, with the step it is waiting for, never zero.
Not stated. A size or licence reads not stated until it is recorded from the model's own source.
Status, accuracy and resources. Read from the same release record as the speeds: the status from the newest published release, accuracy and resources from the release that measured the model. A value no published release recorded reads Results pending.
Which serve. Each figure is measured on Semurg's own serve, which reads the model's weights from disk.

Generated from the release record: models_catalog.json 324f49784f00, release_series.json acbae2415143 and release_baseline.json f5a9b57f4241 (the first 12 characters of each file's SHA-256).

The four operations at every release Memory for AI agents