MiniCPM-V-4.7-35B-A3B-MNN (text only)

Published by TokForge for on-device inference.

TokForge

• Website: https://tokforge.ai
• Discord: https://discord.gg/Acv3CBtfVm
• Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
• iOS TestFlight: https://testflight.apple.com/join/jnufjzRr

Runs on-device in the TokForge app.

Text only

This bundle carries the language model only. The MiniCPM-V vision encoder is not included, because TokForge's engine has no adapter for it yet. In TokForge, photos are described by the app's built-in image helper instead, and the model answers from that description.

What this is

MiniCPM-V 4.7 35B-A3B by OpenBMB, in the 35B Mixture-of-Experts size (35B total, about 3B active per token). Its language model follows the Qwen3.6 Mixture-of-Experts design: 40 layers, 256 experts with 8 active per token, hidden size 2048. This repo holds that language model converted to MNN for phones: 4-bit weights and BF16 embeddings, 22.5 GB of model files. The graph uses the layout that the MNN engine in TokForge 1.3.8.1 and 1.3.9.4 loads. In the TokForge catalog it is listed for phones with 24 GB of RAM or more.

Files

file	bytes	sha256
config.json	502	801311d0461dceffb4fe06f71a38b8ffb9616022626109537d3343ecc747619f
llm_config.json	7,786	6f169c52e203dd336553c73cc1dcb7dcafb97fd40e69068a37738e5e78130511
llm.mnn	54,337,952	1cd28f1170a1ca2f57ca618393444ef3d7d24003bd30ab3671f6184c916e7211
llm.mnn.json	119,473,194	16241d7e7963f0f66610e3069ebab882677beffecef54ad63ec14f29fdec5bd2
llm.mnn.weight	21,345,939,870	099fbb2a5e818832166b3ee65db93a4899cbf018ac0ed281029fe2b7118e27bc
embeddings_bf16.bin	1,016,397,824	4fc37f2361f58ecc2e273ef9a584ed7273388e7a4de7868cd409e6d3ae8ddfe2
tokenizer.mtok	7,336,304	2ea7ed21e8308a0a958d8382a2dcfb71429d5a47c58082225c3ddca6fc42b5c8
export_args.json	1,154	ce06a433d2078d3315c3e84d5388f507f5464028a315d554124a61cdfb2f4dca
LICENSE	11,358	cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30
NOTICE	390	bc8506ed3b8264a57a01df82009f12db6473d2b8d2ddd1059b3c58063b2ec6d4

config.json names tokenizer.mtok (MNN's binary tokenizer). llm.mnn.json is the readable form of the graph. SHA256SUMS lists the same hashes.

How it was made

• Source: openbmb/MiniCPM-V-4.7-35B-A3B at commit 896c2d454672476a7b45dcc5e3f0c700a0218577.
• Converted with MNN's llmexport (TokForge's pinned build): --quant_bit 4 --quant_block 64 --lm_quant_bit 4 --lm_quant_block 64 --hqq --seperate_embed. The exporter change mapped the minicpmv4_7 text backbone onto the existing Qwen3.5 text path; the vision tower and its merger are not exported.
• Linear layers, the output head included: 4-bit HQQ, block size 64. Embeddings: BF16, in embeddings_bf16.bin.
• Graph: the plain LinearAttention layout (no transformer_c4 regions, no folded gate). llm_config.json declares model_type: qwen3_5_moe, the language model's own type.
• Chat template: OpenBMB's own template, byte for byte (thinking is off unless the request turns it on).

Accuracy on the host

Measured on an x86 build of the MNN engine source that TokForge ships (CPU, with this bundle's own settings), against the original model in float32. Three long real-text prompts (453, 912 and 1,360 tokens) in the model's chat template, scored on the first reply token and 16 more teacher-forced steps each: 51 positions in total.

measure	result
first reply token matches the original	3 of 3
top-1 token matches the original	48 of 51
KL divergence to the original, median / mean	0.0014 / 0.060

Where the top token differs, both models pick another valid word (for example "outlines" against "permissions" after "This text defines key terms and").

License

Apache-2.0, as the original model. LICENSE holds the license text and NOTICE credits OpenBMB.

Credits

• MiniCPM-V 4.7 35B-A3B by OpenBMB (Hugging Face). Used under the Apache-2.0 license.
• Its language backbone builds on Qwen3.5 and Qwen3.6 by the Qwen team (Alibaba Cloud), as OpenBMB credits them (Qwen3.5, Qwen3.6). Apache-2.0.
• Converted with llmexport from MNN by Alibaba (Apache-2.0).