Instructions to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Use Docker

docker model run hf.co/gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

LM Studio
Jan

vLLM

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Ollama
How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with Ollama:
```
ollama run hf.co/gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
```
Unsloth Desktop

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with Docker Model Runner:
```
docker model run hf.co/gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL
```

Lemonade

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Run and chat with the model

lemonade run user.MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE:UD-Q4_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

MiniCPM-V-4.7-35B-A3B Uncensored — Sangreal ICE

Five ICE 2.0 tiers of openbmb's MiniCPM-V 4.7 35B-A3B (35B total, 3B active, Qwen3.5-MoE text tower + SigLIP2 vision), uncensored trunk (openbmb's BF16 + Atomic Chat's refusal-direction LoRA, merged), quantized against Sangreal — a purpose-built 77-bucket calibration corpus — and shipping Qwen-Sharp v22.5.0-arrow4.0-minicpm — our agent-tuned chat template, a heavily modified derivative of Qwen-Sharp, fitted to MiniCPM-V's multimodal content format (spec below). Vision works through openbmb's own projector, shipped here unmodified. Sangreal-UD — Unsloth Dynamic 2.0 replays (with Sangreal10K and default chat-template) — on the same trunk are shipped as well.

⚠️ Uncensored. Refusal behaviour has been suppressed in the weights. Sandbox it at the OS level and control its network and code-execution access; with no refusal backstop, a prompt injection from a hostile page or third-party code has nothing to stop it.

ICE 2.0

All five tiers ship as ICEv2: 12G-ICEv2, 15G-ICEv2, 19G-ICEv2, 23G-ICEv2 and 25G-ICEv2.

Every tier is an ICE quantization of this trunk — Sangreal importance matrix, ICE recipe — finished by a Sangreal-calibrated post-quantization repair of the routed experts, fitted against this trunk's own BF16. Same GGUF types, stock llama.cpp.

The trunk

BF16/ (69.4 GB) = openbmb's official BF16 + Atomic Chat's rank-1 refusal-direction LoRA, merged at scale 1.0 (f32, one BF16 rounding).

Edit: 120 residual writers (attention / linear-attention outputs, every routed and shared expert's down projection). Router, norms, embedding and lm_head untouched.
Verified: quantized to Q8_0, all 312 Q8_0 tensors are byte-identical to Atomic Chat's own baked build.
Cost: 0.0090 mean KLD to the original BF16; MMLU 84.25 → 84.50 (Atomic Chat's measurements).
Imatrix: Sangreal 10K, measured on a Qwen3.6-35B-A3B sibling. On the same 400 chunks it differs from MiniCPM's own by about as much as two corpus slices differ (routing TVD 0.108 vs 0.092).

At a glance — against Sangreal-UD at the same size

Mean KLD to BF16, one session. Size-matched = the Unsloth Dynamic 2.0 recipes (UD-Q2_K_XL, UD-IQ3_XXS, UD-IQ4_XS, UD-Q4_K_XL, UD-Q5_K_XL) replayed tensor-for-tensor on this same trunk with the same importance matrix, interpolated (log-linear) at each ICE file's exact byte size; negative = ICE closer to BF16. The recipe is the only difference. Paired = the same 64 chunks scored for both files against the same reference.

tier	size	code @2K	code @16K	wiki @2K	paired, code @2K (t, chunks ICE better)
12G-ICEv2	11.90 GB	+17.5% †	+19.3% †	+30.7% †	vs Q2_K_XL (12.29 GB): +17.5% (t +5.4, 7/64)
15G-ICEv2	14.51 GB	+2.5%	+1.1%	+6.7%	vs IQ3_XXS (13.21 GB): -25.8% (t -10.4, 63/64); vs IQ4_XS (17.73 GB): +127.8% (t +7.8, 0/64)
19G-ICEv2	18.47 GB	-14.7%	-15.9%	-5.4%	vs IQ4_XS (17.73 GB): -26.9% (t -10.6, 64/64); vs Q4_K_XL (22.36 GB): +92.6% (t +7.5, 1/64)
23G-ICEv2	22.52 GB	-3.5%	-4.3%	+4.1%	vs Q4_K_XL (22.36 GB): -6.1% (t -3.4, 43/64); vs Q5_K_XL (26.59 GB): +88.6% (t +7.6, 1/64)
25G-ICEv2	24.53 GB	-13.2%	-11.0%	-4.3%	vs Q4_K_XL (22.36 GB): -39.3% (t -7.8, 63/64); vs Q5_K_XL (26.59 GB): +21.8% (t +4.9, 6/64)

Why a direct comparison is hard here. Unsloth Dynamic 2.0's ladder and the ICE tiers sit at different sizes (gap = UD file − ICE file):

ICE tier	size	nearest Sangreal-UD	gap	comparison
12G-ICEv2	11.90 GB	UD-Q2_K_XL 12.29 GB	+0.39 GB	direct; UD has nothing smaller
15G-ICEv2	14.51 GB	UD-IQ3_XXS 13.21 / UD-IQ4_XS 17.73 GB	−1.30 / +3.22 GB	interpolated across 4.5 GB (UD-Q3_K_XL, ~16.8 GB, not replayed)
19G-ICEv2	18.47 GB	UD-IQ4_XS 17.73 GB	−0.74 GB	direct: ICE −26.9%, 64/64 chunks
23G-ICEv2	22.52 GB	UD-Q4_K_XL 22.36 GB	−0.16 GB	direct: ICE −6.1%, 43/64 chunks
25G-ICEv2	24.53 GB	UD-Q4_K_XL 22.36 / UD-Q5_K_XL 26.59 GB	−2.17 / +2.06 GB	interpolated across 4.2 GB (UD-Q4_K_S, ~20.9 GB, not replayed)

Where a UD tier sits within 0.8 GB below an ICE tier (19G, 23G), ICE wins the paired comparison on the same chunks (code @2K). 12G is 0.39 GB smaller than the only nearby UD file and loses to it; the 15G and 25G rows are interpolations, not head-to-heads.

† 12G-ICEv2 is outside Unsloth Dynamic 2.0's ladder, so instead of a size-matched point it is compared directly against Q2_K_XL (12.29 GB, 0.39 GB larger).

† 19G-ICEv2, 23G-ICEv2 and 25G-ICEv2 are clear winners in their respective category.

Code at 16K context (code.test.raw, n_ctx 16384 × 8), mean KLD to BF16 vs file size; lower-left is better. All points from one session (one base, one binary); raw numbers in measurements/2026-10-14/ and Table 2.

Measurements

One binary (llama.cpp 3653e6d), one reference (this trunk's own BF16, in this repo), one session (2026-10-14, 8× RTX 4090). 64 chunks at n_ctx 2048; the long-context table uses 8 chunks at n_ctx 16384. Every file below was measured in this session. Raw logs: measurements/2026-10-14/.

Table 1 — Code (`code.test.raw`)

PPL(base) 2.227770 · BF16 = 100

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 26.6 GB
Sangreal-UD-Q5_K_XL	26.59 GB	0.004029	0.0618	0.3285	1.0022	98.51 %	7.833	6.138	99.3
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.004910	0.0755	0.3833	1.0015	98.37 %	7.386	5.662	99.2
≈ 22–23 GB
Sangreal 23G-ICEv2	22.52 GB	0.007601	0.1217	0.6022	1.0026	98.01 %	7.215	5.197	98.9
Sangreal-UD-Q4_K_XL	22.36 GB	0.008086	0.1318	0.6784	1.0032	97.98 %	7.473	5.161	98.8
≈ 18 GB
Sangreal 19G-ICEv2	18.47 GB	0.015575	0.2640	1.2559	1.0092	97.25 %	6.872	4.263	98.1
Sangreal-UD-IQ4_XS	17.73 GB	0.021321	0.3603	1.4235	1.0101	96.66 %	6.757	4.092	97.5
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.048564	0.8802	3.5920	1.0335	95.09 %	6.531	3.350	95.3
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.101390	1.8525	6.0519	1.0780	92.91 %	6.313	2.746	91.4
Sangreal-UD-IQ3_XXS	13.21 GB	0.065436	1.2006	4.1866	1.0398	94.22 %	5.474	3.049	94.0
Sangreal-UD-Q2_K_XL	12.29 GB	0.086275	1.5367	4.9039	1.0578	93.15 %	4.700	2.837	92.4

Table 2 — Code at long context (`code.test.raw`, n_ctx 16384 × 8)

PPL(base) 1.960705 · BF16 = 100

Compare rows within this table only.

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 26.6 GB
Sangreal-UD-Q5_K_XL	26.59 GB	0.003083	0.0509	0.2609	1.0011	98.80 %	7.833	6.138	99.4
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.003963	0.0614	0.3832	1.0012	98.63 %	7.386	5.662	99.3
≈ 22–23 GB
Sangreal 23G-ICEv2	22.52 GB	0.006098	0.0965	0.5830	1.0024	98.29 %	7.215	5.197	99.1
Sangreal-UD-Q4_K_XL	22.36 GB	0.006554	0.1135	0.6148	1.0028	98.29 %	7.473	5.161	99.0
≈ 18 GB
Sangreal 19G-ICEv2	18.47 GB	0.012302	0.2121	1.1660	1.0071	97.60 %	6.872	4.263	98.4
Sangreal-UD-IQ4_XS	17.73 GB	0.017039	0.2897	1.4315	1.0076	97.18 %	6.757	4.092	98.0
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.038140	0.6793	3.5125	1.0288	95.87 %	6.531	3.350	96.2
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.081199	1.5631	5.7263	1.0658	93.93 %	6.313	2.746	92.9
Sangreal-UD-IQ3_XXS	13.21 GB	0.052030	0.9736	3.9749	1.0343	95.17 %	5.474	3.049	95.1
Sangreal-UD-Q2_K_XL	12.29 GB	0.068050	1.2428	4.6511	1.0470	94.28 %	4.700	2.837	93.8

Table 3 — English text (WikiText-2)

PPL(base) 5.981634 · BF16 = 100

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 26.6 GB
Sangreal-UD-Q5_K_XL	26.59 GB	0.006247	0.0751	0.2920	1.0000	97.03 %	7.833	6.138	98.7
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.008174	0.1004	0.4065	1.0019	96.44 %	7.386	5.662	98.4
≈ 22–23 GB
Sangreal 23G-ICEv2	22.52 GB	0.012065	0.1546	0.5345	1.0025	95.75 %	7.215	5.197	97.9
Sangreal-UD-Q4_K_XL	22.36 GB	0.011869	0.1450	0.6107	1.0046	95.84 %	7.473	5.161	97.9
≈ 18 GB
Sangreal 19G-ICEv2	18.47 GB	0.025328	0.3329	1.1310	1.0116	94.00 %	6.872	4.263	96.5
Sangreal-UD-IQ4_XS	17.73 GB	0.031283	0.3938	1.4568	1.0074	93.23 %	6.757	4.092	95.8
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.071686	0.9510	3.1756	1.0397	90.15 %	6.531	3.350	92.4
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.157371	2.1219	5.4703	1.1064	85.68 %	6.313	2.746	86.2
Sangreal-UD-IQ3_XXS	13.21 GB	0.091588	1.1988	3.6791	1.0434	88.73 %	5.474	3.049	90.7
Sangreal-UD-Q2_K_XL	12.29 GB	0.120373	1.5442	4.7073	1.0644	86.71 %	4.700	2.837	88.5

overall = 0.70/(1 + meanKLD) + 0.30 * sameTop1, ×100. BF16 = 100. The same composite as the Ornith-1.5 ICE card: 70% on how close the whole output distribution stays, 30% on agreement about the argmax.

Read the tail columns as shape, not order. 99.9% KLD is roughly the 33rd-worst token of 32,768 — an extreme order statistic with large sampling variance, so it inverts between adjacent files without meaning anything. Rank on mean KLD.

Two bpw columns. Only 8 of 256 experts fire per token, so a bit in ffn_*_exps is worth ~3% of a bit in attention, the shared expert or the output head. active bpw weights by that (token embedding excluded: it is a lookup, not a matmul); file bpw is just size ÷ parameters.

Don't compare the tables to each other. Code is more predictable text, so every file scores about half the divergence on it. Compare rows within a table.

Every row is measured against this trunk's own BF16, published in this repo. The Unsloth Dynamic 2.0 rows are their recipes applied to the same weights with the same importance matrix, so a difference here is the recipe and nothing else.

What Sangreal is

The calibration corpus these tiers were quantized against. Built from primary sources, not assembled from eaddario's set, bartowski's calibration_datav5, or any other ready-made calibration file. 77 buckets, 9,533 documents, 47,443,549 tokens, sha256 85a6b823a6762bf6….

The full render is 92,663 chunks. These tiers use a 10,000-chunk pass over it (The trunk says which model it was measured on).

block	buckets	share	documents	synthetic
code + cyber + spec	36	45.10%	5,858	0.1%
reasoning	5	13.16%	553	100.0%
agentic	7	11.95%	1,018	65.2%
science + domain	11	11.24%	1,183	0.3%
language	11	9.66%	596	0.0%
mathematics	7	8.89%	325	0.0%
total	77	100%	9,533	21.7%

Split: technical 70.2 · science+domain 20.1 · language 9.7.

Synthetic is 21.7% by bytes and sits in exactly three buckets — teacher:k2-horizon, teacher:deepseek-v4-pro, teacher:hermes. Everything outside reasoning and agentic is real text.

Language — 11 languages, 5 non-Latin scripts

bucket	share	docs	bucket	share	docs
`arabic`	0.92%	48	`russian`	0.87%	58
`chinese`	0.89%	31	`spanish`	0.87%	72
`hindi`	0.88%	44	`portuguese`	0.87%	70
`japanese`	0.88%	33	`turkish`	0.87%	56
`french`	0.88%	60	`german`	0.86%	73
`english`	0.87%	51

The spread across the whole block is 0.06 points, and that is deliberate: language's job here is activation, not ranking. Each bucket sits just above the saturation floor at 44.9% consumption, so these buckets select for the first time — the legacy v1 spec had consumed 96% of turkish and 92% of english.

Mathematics — absent from the pool until this cycle

bucket	share	docs
`math` (general)	3.16%	170
`analysis`	1.24%	28
`algebra`	1.16%	46
`geometry`	1.12%	30
`proof`	1.08%	28
`topology`	1.05%	20
`number-theory`	0.06%	3

Sourced from the LaTeX algebraic geometry, mathlib4, UniMath, math-comp, open-web-math, AutoMathText and proof-pile.

The imatrix and the recipes are not published with this release.

What's inside

ICE = Isolation of Compounding Error: allocate bits by how far a quantization error travels, not by how large the activations are. Routers stay F32, SSM decay gates stay F32, the KV-cached projections stay F16, the always-on dense path stays Q8_0, and the budget goes to the routed expert bank — most of the parameters, but only 8-of-256 active per token. ICE 1.6 keeps token_embd at Q8_0 (its error propagates with context) and moves the output head to Q6_K (its error stays flat from 2K to 16K). Method and the cases where it does not win: gbuzhf/ICE-quantization.

No MTP head. The config declares one MTP layer, but the checkpoint carries no MTP weights. These tiers are the 40-block trunk; there is no built-in draft model. A straight conversion copies mtp_num_hidden_layers into block_count (and a 41-entry recurrent_layers), producing a GGUF that advertises 41 blocks and looks for a block that does not exist. Every file in this repo declares 40, verified on each one.

Chat template: Qwen-Sharp v22.5.0-arrow4.0-minicpm (ours; spec below). MiniCPM-V 4.7 keeps the Qwen3.5/3.6 chat tokens, so the template's delimiters are native tokens — checked on every tier at signing.

Vision. The edit touches only the language model's residual writers, so the projector is openbmb's own, shipped byte-for-byte: mmproj-MiniCPM-V-4.7-35B-A3B-F16.gguf (default) and mmproj-MiniCPM-V-4.7-35B-A3B-F16-4x.gguf (4× visual-token compression). Load with --mmproj <file>; MiniCPM-V 4.7 needs llama.cpp 69f201a or later.

Chat template — Qwen-Sharp v22.5.0-arrow4.0-minicpm

Every ICE GGUF embeds arrow4.0-minicpm (the two UD files keep openbmb's own template); to pin it explicitly, pass the file shipped here:

llama-server -m <model>.gguf --jinja --chat-template-file chat_template_sharp-v22.5.0-arrow4.0-minicpm.jinja

arrow4.0 is our agent layer on Qwen-Sharp v22.5.0. Plain chat stays byte-identical to Sharp; with tools it adds short agent rules, one note per tool result, a STATUS line, thinking only at decision points, .arrow project memory and a hand-back report when work stalls. The -minicpm file renders images, videos and multi-part content exactly as openbmb's own template does (a bare <|image_pad|>, parts joined by a newline); text, tools and thinking are arrow4.0's.

Full description and switches: gbuzhf/Qwen-Sharp-Arrow-Chat-Templates.

Files

file	size	sha256
`MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-12G-ICEv2.gguf`	11.90 GB	`e7c4c1799571…`
`MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-15G-ICEv2.gguf`	14.51 GB	`0f276dc0f05f…`
`MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-19G-ICEv2.gguf`	18.47 GB	`e3bceae66ee1…`
`MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-23G-ICEv2.gguf`	22.52 GB	`c2afa57d9ff4…`
`MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-25G-ICEv2.gguf`	24.53 GB	`ba1f4af5af73…`
`UD2/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-UD-Q2_K_XL.gguf`	12.29 GB	`d47aab169327…`
`UD2/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-UD-IQ4_XS.gguf`	17.73 GB	`e60d066b39e1…`
`mmproj-MiniCPM-V-4.7-35B-A3B-Q8_0.gguf`	0.73 GB	`7c837b42a485…`
`mmproj-MiniCPM-V-4.7-35B-A3B-F16.gguf`	1.12 GB	`bff3b688e14c…`
`mmproj-MiniCPM-V-4.7-35B-A3B-F16-4x.gguf`	0.91 GB	`30f4daac0f61…`
`MiniCPM-V-4.7-35B-A3B-Uncensored-Q8_0.gguf`	36.90 GB	`e1e2b777cced…`
`BF16/MiniCPM-V-4.7-35B-A3B-Uncensored-BF16-00001-of-00002.gguf`	44.87 GB	`b82fc991c265…`
`BF16/MiniCPM-V-4.7-35B-A3B-Uncensored-BF16-00002-of-00002.gguf`	24.51 GB	`86525e24f1f9…`

measurements/2026-10-14/ — raw llama-perplexity output for every row of every table (ICEv2 and the Unsloth Dynamic 2.0 ladder, × 3 columns) and the three BF16 base runs.

Credits


base model	openbmb/MiniCPM-V-4.7-35B-A3B
refusal-direction LoRA	AtomicChat/MiniCPM-V-4.7-35B-A3B-Uncensored-GGUF
vision projector	openbmb/MiniCPM-V-4.7-35B-A3B-gguf
comparison recipes	Unsloth Dynamic 2.0
chat template	peculiar-ragdoll/Qwen-Sharp-Chat-Templates
ICE method	gbuzhf/ICE-quantization

KLD measures fidelity to this repo's reference and nothing else — not reasoning, tool use or speed. It is also within-model: every file here is measured against this trunk's own BF16, so these values are not comparable to another repo's. Compare across models with PPL.

Downloads last month: -

GGUF

Model size

35B params

Architecture

qwen35moe

Hardware compatibility

2-bit

3-bit

4-bit

5-bit

8-bit

16-bit

View +5 variants

Model tree for gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE

Base model

openbmb/MiniCPM-V-4.7-35B-A3B

Quantized

(47)

this model

Collection including gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE

ICE 2.0

Collection

2 items • Updated about 19 hours ago

MiniCPM-V-4.7-35B-A3B Uncensored — Sangreal ICE

ICE 2.0

The trunk

At a glance — against Sangreal-UD at the same size

Measurements

Table 1 — Code (code.test.raw)

Table 2 — Code at long context (code.test.raw, n_ctx 16384 × 8)

Table 3 — English text (WikiText-2)

What Sangreal is

Language — 11 languages, 5 non-Latin scripts

Mathematics — absent from the pool until this cycle

What's inside

Chat template — Qwen-Sharp v22.5.0-arrow4.0-minicpm

Files

Credits

Model tree for gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE

Collection including gbuzhf/MiniCPM-V-4.7-35B-A3B-Uncensored-Sangreal-ICE

Table 1 — Code (`code.test.raw`)

Table 2 — Code at long context (`code.test.raw`, n_ctx 16384 × 8)