license: apache-2.0 base_model:

• openbmb/MiniCPM-V-4.7-35B-A3B base_model_relation: quantized pipeline_tag: text-generation library_name: gguf language:
• en
• zh tags:
• gguf
• ICE
• ice-quant
• imatrix
• sangreal
• moe
• qwen35moe
• minicpm-v
• vision
• abliterated
• uncensored

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 (one refusal direction projected out of the official BF16), 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. Unsloth Dynamic 2.0 replays (with Sangreal10K and default chat-template) on the same trunk are shipped as well.

⚠️ Uncensored. This is an abliterated checkpoint — 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

Our own BF16 master, published in BF16/: openbmb's official BF16 weights with one refusal direction projected out — W' = W − r rᵀW on the 120 matrices that write the residual stream (40 attention / linear- attention outputs, the down projections of all 256 routed experts in every block, 40 shared experts), applied from the published rank-1 adapter at scale 1.0 in f32 and rounded to BF16 once. The router, every input projection, every norm, the embedding and lm_head are untouched. We verified the bake instead of trusting it: quantized to Q8_0 (which ignores the importance matrix, so its bytes are reproducible), all 312 Q8_0 tensors of our master — including the 80 edited non-expert writers — are byte-identical to the adapter author's own baked build, and every edited writer keeps at most 7.6e-3 of its component along r after rounding.

The importance matrix is Sangreal's 10,000-chunk pass measured on a Qwen3.6-35B-A3B fine-tune of the same architecture. On a sparse MoE an imatrix is mostly a statement about which experts fire, so we chose it by the router: MiniCPM-V 4.7's own vision tune moved its router 7.5–9.2% (relative L2, layers 0–39) away from Qwen3.6, and the matrix's source sits at the same distance (7.6–9.4%) — the closest Sangreal matrix there is.

At a glance — against Unsloth Dynamic 2.0 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.

† 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.

Mean KLD vs file size, code at 16K context: Sangreal ICE 2.0 and Unsloth Dynamic 2.0's own GGUF ladder

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

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.

Table 3 — English text (WikiText-2)

PPL(base) 5.981634 · BF16 = 100

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).

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

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

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

• 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

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.