radiosilent/vectorstream-mbestudio-sysml-assistant

MBE-30B

MBE-30B is the 30B local model that MBEstudio offers for writing SysML v2 models. It runs on a single 16 GB GPU through Ollama, and it is the model the studio's generation loop is tuned around for local use.

What this release is

This first release is not fine-tuned. It is Qwen3-Coder-30B-A3B-Instruct (30.5B parameters, 3.3B active, mixture of experts) at Q4_K_M, packaged with:

• a SysML v2 system prompt for MBEstudio's profile and the SysML v2 standard library,
• a 32k context window by default (the studio runs it at 64k) and sampling settings that match the studio.

We picked this base by evaluation. A fine-tuned release will replace these weights only if it beats this base on the same evaluation.

Evaluation

The evaluation runs MBEstudio's full generation loop: plan, write each package, compile with the MBEstudio SysML v2 compiler, and repair from the compiler's diagnostics. It runs at a 32k context window. Each brief is held out from the few-shot examples.

Brief	Packages	Clean on first pass	Repairs	Errors at the end
Commuter EV	3	0 of 3	2	0
L4 robotaxi	3	1 of 3	2	0
Island dredger	3	0 of 3	2	0
Midsize aircraft	3	1 of 3	9	4 (at 64k; see below)

The first three briefs ended with zero compiler errors. The midsize aircraft is the largest brief: at a 32k context its parts package did not fit, so it was run at 64k, the context the studio uses for this model. It produced all three packages from the model and ended with 4 errors, all M0604: three envelope dimensions defined in terms of themselves, which the repair loop did not resolve in four attempts. That run took about 27 minutes on an RTX 5080. In an earlier run of the same evaluation, qwen3:30b, gpt-oss:20b and qwen3:14b finished 0 of 4 briefs clean.

Most first-pass errors are MBEstudio profile rules, such as metadata values written with units and imports without visibility. The repair loop fixes them. The model needs that compile-and-repair loop. Used on its own, without a compiler in the loop, expect output that does not compile.

Use with Ollama

hf download radiosilent/vectorstream-mbestudio-sysml-assistant mbe-30b-Q4_K_M.gguf Modelfile --local-dir mbe-30b
cd mbe-30b
ollama create mbe-30b -f Modelfile
ollama run mbe-30b

The Modelfile uses Ollama's built-in Qwen3-Coder renderer (RENDERER/PARSER qwen3-coder); the GGUF's own chat template does not continue an assistant prefill, which MBEstudio uses to start each code block. About 19 GB on disk. On a 16 GB GPU, part of the model runs on the CPU. Expect roughly 60 tokens/s on an RTX 5080 at a 24k to 32k context.

Use in MBEstudio

Pick mbe-30b in the studio's model menu when the studio runs against local Ollama.

License

Apache 2.0, the license of the Qwen3-Coder base model. The weights are the Q4_K_M quantization published by Unsloth, unchanged.