pd95/ apertus-mlx:8b

331 4 weeks ago

Experimental MLX-backed Apertus 8B Instruct for custom Ollama builds. Requires ApertusForCausalLM support and is not compatible with the regular Ollama app.

tools 8b 
ollama run pd95/apertus-mlx:8b

Details

4 weeks ago

a9e7563fb7cf · 5.3GB

json
{ "_from_model_config": true, "bos_token_id": 1, "eos_token_id": [2, 68, 72], "transformers_version"
json
{ "bos_token": { "content": "<s>", "lstrip": false, "normalized": false, "rstrip": false, "single_wo
json
{ "version": "1.0", "truncation": null, "padding": null, "added_tokens": [ { "id": 0, "content": "<u
json
{ "add_bos_token": true, "add_eos_token": false, "add_prefix_space": false, "added_tokens_decoder":
license
Copyright 2025 - The Swiss AI team. Apache License Version 2.0, January 2004 http://www.apache.org/l
params
{ "stop": [ "</s>", "<|assistant_end|>", "<|tools_suffix|>" ], "
json
{ "architectures": [ "ApertusForCausalLM" ], "attention_bias": false, "attention_dropout": 0.0, "bos
template
{{ .Prompt }}
451 tensors

Readme

apertus-mlx

Experimental MLX-backed Ollama build of Apertus 8B Instruct.

This model requires a custom Ollama build with experimental MLX safetensors support and ApertusForCausalLM support. It will not run on the regular public Ollama app/build.

The artifacts require an Apertus-capable custom MLX preview. Their stored numeric compatibility requirement is Ollama 0.25.0-rc0. Use the v0.34.1 MLX preview or newer for the current packed-MXFP8 artifact, correct capability reporting, and rejection of unsupported thinking requests.

If you are ready to install a custom Ollama build, visit https://pd95.github.io/Ollama/.

Tags

  • pd95/apertus-mlx:8b — recommended default, currently NVFP4
  • pd95/apertus-mlx:8b-nvfp4 — explicit NVFP4 artifact
  • pd95/apertus-mlx:8b-mxfp8 — larger packed-MXFP8 artifact
  • pd95/apertus-mlx:8b-bf16 — unquantized BF16 artifact

Immutable release tags:

  • pd95/apertus-mlx:8b-nvfp4-v0.32.15-r2
  • pd95/apertus-mlx:8b-mxfp8-v0.34.1 — current packed-MXFP8 artifact
  • pd95/apertus-mlx:8b-mxfp8-v0.32.15-r2
  • pd95/apertus-mlx:8b-bf16-v0.32.15-r2

Source

This model is an experimental Ollama/MLX conversion of:

https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509/tree/50761a511195fde9d958f62f3b6344329d4bd191

Apertus 8B Instruct is released by the Swiss AI Initiative / Swiss National AI Institute.

Technical report:

https://arxiv.org/abs/2509.14233

License and use policy

Apertus is released under the Apache License 2.0:

https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509/blob/main/LICENSE.txt

Use is also subject to the Apertus LLM Acceptable Use Policy:

https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509/blob/main/USAGE_POLICY.md

This repository only republishes a quantized MLX/Ollama artifact. It does not change the upstream Apertus license or usage terms. Please review the upstream Apache-2.0 license and Apertus acceptable-use policy before use or redistribution.

Notes

  • Architecture: Apertus 8B Instruct
  • Source revision: 50761a511195fde9d958f62f3b6344329d4bd191
  • Available formats: NVFP4, packed MXFP8, and BF16
  • Context window: 64K
  • Capabilities: text, tools
  • Thinking: not supported by Apertus 1.0
  • Runtime: experimental Ollama MLX runner
  • Current MXFP8 import preview: v0.34.1
  • Stored execution requirement: v0.25.0-rc0
  • Current MXFP8 validation runtime: v0.34.1
  • Recommended MLX preview: v0.34.1 or newer

The v0.32.15-r2 artifacts corrected the stored Apertus family and capability metadata. In particular, they no longer advertise thinking support.

The current 8b-mxfp8 alias and immutable 8b-mxfp8-v0.34.1 tag additionally use MLX’s corrected MXFP8 E8M0 scale selection. Block scales are rounded upward so the largest E4M3 values do not saturate. The change affects the six projection matrices in each of 32 layers plus lm_head.weight; deterministic inference, Apertus tool behavior, quality probes, performance, and memory were validated before publication. The previous 8b-mxfp8-v0.32.15-r2 tag remains available as the verified rollback artifact.

The retained v0.32.15-r2 artifact was execution-tested back to the first Apertus preview. The current MXFP8 artifact uses the same packed format and stores the same requirement, but this closeout validated its changed tensor generation on v0.34.1; it did not rerun the new artifact across every older runtime. Older previews may also dynamically report duplicate tool support and the unsupported thinking capability.

This model is intended for testing the experimental Apertus MLX path in Ollama.