pd95/ apertus-mini-mlx:4b-nvfp4

419 1 week ago

Experimental MLX-backed Apertus v1.1 Mini Instruct models for custom Ollama builds. Known issue: the current Apertus v1.1 Mini artifacts can produce nonsensical chat

tools thinking 0.5b 1.5b 4b 
ollama run pd95/apertus-mini-mlx:4b-nvfp4

Details

1 week ago

69d822b0970b · 2.7GB

template
{{ .Prompt }}
json
{ "architectures": [ "ApertusForCausalLM" ], "attention_bias": false, "attention_dropout": 0.0, "bos
json
{ "_from_model_config": true, "bos_token_id": 1, "eos_token_id": [ 68 ], "pad_token_id": 10, "transf
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
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "stop": [ "</s>", "<SPECIAL_68>", "<SPECIAL_72>" ], "temperatu
339 tensors

Readme

apertus-mini-mlx

Experimental MLX-backed Ollama builds of the Apertus v1.1 Mini Instruct models, available in 0.5B, 1.5B, and 4B parameter sizes.

These models require a custom Ollama build with experimental MLX safetensors support and ApertusForCausalLM support. They will not run on the regular public Ollama app/build.

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

Known limitation

The currently published Mini artifacts can produce nonsensical chat responses and may disregard information supplied earlier in a conversation. This affects the 0.5B, 1.5B, and 4B BF16 and NVFP4 variants. They are not recommended for conversational use while the issue is investigated. Prompt rendering appears correct; the cause may be upstream Mini model quality or a Mini-specific MLX runtime parity issue. Apertus 1.0 8B is not affected.

Tags

  • pd95/apertus-mini-mlx:0.5b, :1.5b, and :4b — current NVFP4 defaults; see the known limitation above
  • pd95/apertus-mini-mlx:<size>-nvfp4 — explicit NVFP4 artifacts
  • pd95/apertus-mini-mlx:<size>-bf16 — explicit BF16 artifacts
  • pd95/apertus-mini-mlx:<size>-<quantization>-v0.32.15-r2 — immutable release tags

The six initial immutable artifacts are 0.5b-bf16-v0.32.15-r2, 0.5b-nvfp4-v0.32.15-r2, 1.5b-bf16-v0.32.15-r2, 1.5b-nvfp4-v0.32.15-r2, 4b-bf16-v0.32.15-r2, and 4b-nvfp4-v0.32.15-r2.

Source

These models are experimental Ollama/MLX conversions of Apertus v1.1 Mini Instruct checkpoints:

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

License and Use Policy

The upstream Apertus v1.1 Mini checkpoints are released under the Apache License 2.0. Each published artifact includes an Apache-2.0 license layer.

This repository republishes converted or quantized MLX/Ollama artifacts. It does not change the upstream license or usage terms. Please review the upstream model card and license before use or redistribution.

Notes

  • Architecture: Apertus v1.1 Mini Instruct
  • Quantizations: BF16 and NVFP4
  • Stored capabilities: text, tools, thinking; these do not establish conversational reliability
  • Minimum supported custom MLX preview: v0.32.6
  • Validated custom MLX preview range: v0.32.6 through v0.32.15-r2
  • Runtime: experimental Ollama MLX runner

These models are retained for experimental investigation of the Apertus MLX path in Ollama. They are not supported by the regular public Ollama app/build.