Runtime: Osaurus 0.25.21 or newer (
osaurus.json:required_osaurus_version).
OsaurusAI/MiniCPM-V-4.7-1B-JANG_8M
A JANG bundle of openbmb/MiniCPM-V-4.7-1B: a dense 1B model (24 layers, GatedDeltaNet linear attention and full attention 3:1), with image, multi-image and video understanding, thinking on by default, and Qwen-style XML tool calls. 3.0 GiB of bf16 weights in 2.0 GiB.
What is quantized how
| part | format |
|---|---|
| MLP gate / up / down | affine 8-bit, group 64, GPTQ |
| attention q/k/v/o, GatedDeltaNet projections | affine 8-bit, GPTQ |
| embeddings, lm_head | affine 8-bit |
| vision tower + projector | bf16 (kept at source precision) |
| norms, conv, A_log, dt_bias | source values |
Every quantized matrix with an input Hessian was rounded with GPTQ. The Hessians come from the model's own outputs: the bf16 model wrote reasoning traces (thinking on and off), tool calls, code, long-document answers and multi-turn agent turns across 11 prompt families at its native sampling settings, plus 192 images and videos through the vision path, and those sequences were the calibration data. Affine quantization is MLX's own grid (bf16 scales and biases), so the bundle loads at full speed on the standard kernels.
Fidelity vs the bf16 model
126,332 teacher-forced positions on 164 held-out sequences the bf16 model generated itself (never seen in calibration), scored against the bf16 model: top-128 KL, top-1 agreement, and the change in log-likelihood of the real next token (dNLL).
| size | KL median | KL mean | top-1 | dNLL | |
|---|---|---|---|---|---|
| bf16 model run in bf16 (the floor) | 3.0 GiB | 4.9e-06 | 0.0005 | 99.24% | +0.0003 |
| 1B-JANG_8M | 2.0 GiB | 9.5e-06 | 0.0008 | 99.07% | +0.0006 |
By prompt family (KL median, top-1):
| family | KL median | top-1 |
|---|---|---|
| agentic / tool use | 0.0000 | 98.8% |
| coding | 0.0003 | 98.4% |
| creative writing | 0.0009 | 97.7% |
| cybersecurity | 0.0005 | 98.3% |
| long documents | 0.0002 | 98.6% |
| math | 0.0000 | 99.4% |
| networking | 0.0000 | 99.2% |
| office tasks | 0.0000 | 99.5% |
| science | 0.0000 | 99.1% |
| terminal agent | 0.0000 | 99.0% |
| turn after a tool result | 0.0000 | 99.0% |
Decisions: the quantized model picks the same token as bf16 at 78 / 78 </think> closes and at 29 / 33 tool-call starts (bf16 vs itself: 29 / 33).
Images and video. On 48 held-out image and video items (documents, charts, screenshots, code and terminal shots, invoices, real video, screen recordings), teacher-forced on the bf16 model's own answers: KL mean 0.0020, top-1 agreement 98.7%.
Model notes
- Modalities: text, images (multi-image, high-resolution slicing), video. No audio.
downsample_mode16x (default) or 4x selects the visual token budget. - Reasoning: on by default; the chat template opens
<think>.enable_thinking: falseturns it off. There are no reasoning-effort levels. - Tools: Qwen3-Coder-style XML (
<tool_call><function=…><parameter=…>). - Sampling (
generation_config.json, mirrored injang_config.json): temperature 1.0, top-p 0.95, top-k 20, EOS<|im_end|>/<|endoftext|>. The 1B release ships no sampling values; these are the 35B values. - The source config lists one MTP layer, but the checkpoint contains no MTP weights; none is included.
Source
openbmb/MiniCPM-V-4.7-1Brevision984a89aafcd5, Apache-2.0.- Quantization and bundle by Jinho Jang (eric@osaurus.ai), 2026-10-12.