File size: 2,804 Bytes
de31098
 
a55c228
 
de31098
 
 
 
 
 
 
a55c228
 
de31098
 
a55c228
de31098
a55c228
de31098
a55c228
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de31098
 
 
a55c228
de31098
 
 
 
 
 
 
 
 
 
 
 
 
a55c228
de31098
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
---
license: apache-2.0
base_model: Nanbeige/Nanbeige4.2-3B
base_model_relation: quantized
pipeline_tag: text-generation
library_name: coreai
tags:
- coreai
- apple-silicon
- recurrent-transformer
- text-generation
- quantized
- int8
---

# Nanbeige4.2-3B — Core AI INT8

**INT8 weight-quantized Core AI LanguageBundle** converted from [`Nanbeige/Nanbeige4.2-3B`](https://huggingface.co/Nanbeige/Nanbeige4.2-3B), pinned to source revision [`5ff54fb7ed86ce8e216d78bff5417ab9981de3d4`](https://huggingface.co/Nanbeige/Nanbeige4.2-3B/tree/5ff54fb7ed86ce8e216d78bff5417ab9981de3d4).

This is a Core AI `.aimodel` bundle for Apple silicon. It is not a Transformers, GGUF, or MLX checkpoint.

Ported by [Vadim Smirnov (@ukint-vs)](https://github.com/ukint-vs). Source, reproducible conversion recipe, parity gates, and the optimization report are tracked in [john-rocky/coreai-model-zoo#5](https://github.com/john-rocky/coreai-model-zoo/issues/5).

## Format

| Property | Value |
|---|---|
| Published variant | **INT8 only** |
| Body weights | int8, per-block 32 (`int8hu`) |
| Untied LM head | symmetric absmax int8, per-block 32 (`--head-sym`) |
| Activations / KV cache | fp16 |
| Decode input | static sequence length 1 (`--static-ids`) |
| Bundle size | approximately 4.59 GiB |
| Verified context | 4,096 tokens |

The int4 and mixed int4/int8 candidates failed the reasoning quality gate and are intentionally not published.

## Architecture

Nanbeige4.2 is a looped Llama: **22 physical transformer blocks execute twice**, with RMS normalization after each pass and independent KV history. The bundle therefore has **44 executed/cache layers while storing and quantizing one copy of the 22 physical blocks**.

## Bundle

The complete LanguageBundle is at:

```text
gpu-pipelined/
  nanbeige4_2_3b_decode_int8hu_block32_sym_s1/
    metadata.json
    nanbeige4_2_3b_decode_int8hu_block32_sym_s1.aimodel/
    tokenizer/
```

The embedded tokenizer includes the vendor `chat_template.jinja` with thinking enabled and disabled modes. The `.aimodel` producer is `coreai-core 1.0.0b2`.

## Verification

- Official checkpoint vs Core AI authoring: float32 full and cached logits pass `rtol=1e-4`, `atol=1e-4`; identical 32-token greedy continuation.
- Quantized Core AI engine: token-exact vs fp32 on 24-token Paris, 16-token freezing-point, and 64-token decimal-reasoning gates.
- M4 Max Release benchmark, prompt 128 / generation 256 / 3 runs: **47.37 prefill / 46.35 decode tok/s**.
- 4,096-token boundary: **29.83 prefill / 32.80 decode tok/s**, 9.17 GiB peak RSS, zero swaps.
- iPhone `h18p` device acceptance is pending; no iPhone performance claim is made.

## License

The source model and converted weights are distributed under Apache-2.0. Conversion code in `coreai-model-zoo` is BSD-3-Clause.