BlackwellBoy | MiniCPM-V 4.7 Abliterated | EXL3 4.0 bpw

Prepared by BlackwellBoy. One audited rank-one output-projection modification, primary4.0-bpw language-body target, original BF16 vision encoder/mergers and FP16 language embeddings/head. Model artifacts total20,250,182,151bytes (18.86GiB), effective4.600631258bits per original parameter. Text-only calibration; image/video independently tested. No5.0 build was warranted.

Known failure: the official BF16 model, derived BF16 model and EXL3 all call brown “red” in the held-out video. Corrected content score is44/48 for each, not48/48. The extra native-decoder strict color test fails0/4. No newly failing held-out case was introduced. This meets the original bounded nonregression gate, not a zero-error guarantee. See evidence/final-heldout-review-02.json and FINAL_GATE_INTERPRETATION.md.

Source and behavioural result

Official source: openbmb/MiniCPM-V-4.7-35B-A3B@17e651151c161dce119cd0ac57570f60da8a0b40. Derived CID: local:81a27dbf5109a1344e3a4ad2047f049bbe3147e159276a9b68a73c61d4f36e81. Independent audit:120 declared language output-projection edits,1032 unchanged tensors,1152 finite tensors, protected assets identical. One preparation-pair direction at layer20, strength1.0; standard rank-one orthogonalization attributed to Arditi et al., arXiv2406.11717v3. Not a new modification algorithm.

44 original selection cases retained. A separate12-case legitimate-task panel changed from1 unnecessary refusal to0, with12/12 core rubrics retained after quantization. This is small, topic/language-confounded evidence from agent review, not independent human adjudication or universal “uncensoring.” No harmful-compliance optimization. Modification assets are separate from the reusable runtime adapter.

Actual RTX5090 tests

60 final runtime cases completed, zero caps:48 held-out plus12 supplemental retention. Held-out correctness for official/derived/EXL3 is the same:16x nonthinking18/19,4x nonthinking4/5,16x thinking18/19,4x thinking4/5. All four coding conditions executed successfully in a CPU/time/memory/network/filesystem-isolated sandbox. Single-image/OCR/chart/multi-image core answers all correct. Native held-out video has the same inherited brown/red error. Three clean reload cases have exact generated tokens/stops. Three fresh CLI processes passed text, image and native video launch checks.

120 locked teacher-forced targets across16 cases:

Frozen quantization limits KL<=0.10,top1>=90%,PPLratio<=1.10 all pass. Original0.03 maxabsolute BF16/FP16 adapter diagnostic failures remain failures; unchanged-upstream43-module controls and matched-precision evidence separate backend/precision effects from adapter additions.

Sampled whole-device peak20,429MiB during loading; allocator visual-prefill peak19,981,274,624bytes. Fresh-process model load6.74–7.01seconds with potentially warm filesystem. Median sampled-token TTFT/decode bycondition:16x nonthinking0.073s/149.90tok/s (19cases);4x nonthinking0.231s/160.90tok/s (5cases);16x thinking0.077s/138.87tok/s (19cases);4x thinking0.251s/140.90tok/s (5cases). Short mixed prompts and possible cache hits; not universal performance. Full per-case timings/cache counts and sampled-peak limitations are retained.

Tested envelope: Linux/WSL2, actualRTX5090, Python3.12.3, Torch2.10.0+cu128, CUDA12.8, one request,8192 configured context, maxslices1, maxframes3, stackframes1, no speculation. This does not prove maximum8192-token prompts or longer videos. CLI defaultframebudget4 is not a claim that4-frame workloads were tested; use API max_frames=3 for the measured budget.

Video reproducibility

Identical MP4 bytes decode to slightly different RGB pixels on ARM/x86 with the same library versions (maximum2/255). The first final runtime attempt failed exact input hashing and remains preserved. CPU-imported Spark decoded frames reproduce all four original BF16 input hashes exactly; paired tests use those canonical pixels only at the decoder boundary. Production adapter is unchanged and has no fixture dependency. Native decoding was separately exercised on5090. This does not excuse the inherited color-answer error.

Reusable adapter and attribution

Adapter0.1.3 is the separable engineering deliverable: exact architecture registration, nested tensor mapping, protected visual loading, vision bridge, canvas-position integration and generation/cache handling. It supports the pinned official and compatible derived source, with no checkpoint-specific modification assets or machine paths. The official BF16 vision encoder/mergers are reused; ExLlamaV3 runs the language model. No architecture renaming, text-only substitution or new quantization algorithm claim.

ExLlamaV3 commit3a513ac9f91aae777f317ca2341b8e445b62c64e(MIT), Transformers5.18 commit a906d3c4b65095f2308b6a6a193e934d03b8eb5d(Apache-2.0), official OpenBMB implementation1d6b7d68379974af0375cfdb903613d630da34aa(Apache-2.0). Model Apache-2.0 license retained. Original unpublished adapter glue has no new public redistribution grant; publication requires a license decision and authorization. See adapter/CONTRIBUTION.md and UPSTREAM_LOCK.json. Bounded prior-art search establishes no “world first.”

Install and launch

Use adapter/README.md and its exact requirements-tested.txt with the included wheel in a new isolated environment; existing CUDA12.8 compiler/compatible5090driver required. Clean CPU wheel install and actual campaign CUDA runtime are verified; a second fresh CUDA extension build was not repeated. Keep CUDA extension, Triton and HF caches isolated.

export TORCH_EXTENSIONS_DIR="$PWD/.cache/torch-extensions"
export TRITON_CACHE_DIR="$PWD/.cache/triton"
export HF_HOME="$PWD/.cache/hf"
export TORCH_CUDA_ARCH_LIST='12.0+PTX'
export CUDA_HOME=/usr/local/cuda-12.8
unset EXLLAMA_NOCOMPILE
# Activate your pinned environment; MODEL is the downloaded repository directory.
minicpm47-exl3 "$MODEL" messages.json --mode 16x --context 8192 --max-new-tokens 2048
minicpm47-exl3 "$MODEL" messages.json --mode 4x --thinking --context 8192 --max-new-tokens 2048

Messages are the official JSON list, e.g. [{"role":"user","content":[{"type":"text","text":"Explain recursion briefly."}]}]. For images/video add {"type":"image","url":"/path/image.png"} or {"type":"video","url":"/path/video.mp4"} before the text entry. API: MiniCPMRuntime.generate(messages,max_slices=1,max_frames=3,mode="16x",thinking=False,max_new_tokens=2048); close the runtime after use.

Ancillary factual inaccuracies, verbose/format-noncompliant answers and small authored test coverage remain limitations.

Extended evaluation update

The v1 weights are unchanged. Matched greedy SixCat: official BF16 78.33%, derived BF16 81.67%, stock EXL3 80.00%, derived EXL3 80.83% on120 items each, no truncated answers. All sampled thinking-on/off comparisons are now complete; see the full matrix below. Broader legitimate-topic testing finds residual Chinese political evasions in both modes; abliterated describes the audited weight operation, not universal uncensoring.

RTX5090 synthetic text test:261111 input tokens within262144 total capacity; median97.57 decode tokens/s and128.28s first-token latency over three256-token runs; sampled whole-device peak25612MiB. Nine exact-key retrieval checks at64K/128K/256K passed. This is not a general long-document reasoning or multimodal256K certification.

Release packaging: original glue/source-tool redistribution licence awaits the author decision. Upstream licences remain in force. Historical receipts are preserved in the archived source revision. Branding-only edits to receipt prose are recorded with original and updated hashes in RELEASE_MANIFEST.json; they do not change measured results.

Completed v1 comparison matrix

Each SixCat cell covers the same 120 items. Scores are percentages; profiles must not be averaged together.

• One answer reached 32768 tokens in a repetitive instruction loop. These cells are not uncapped successes. † An instruction-loop failure occurred despite EOS. All failures remain included. Greedy budget was 8192; sampled budget 32768 with 49152 context, temperature1/top-k20/top-p.95/seed470. One sampled seed is not an uncertainty estimate. BF16-versus-EXL comparisons include backend/precision differences.

A corrected transport boundary passes genuine direct answers to the unchanged original grader, consistently for all four variants, without changing labels/rubrics or rerunning generation. See comparison details.

The broader completed behavioral review covers 24 legitimate prompts in both thinking modes: stock EXL3 had39/48 substantive answers and9 refusals/evasions; this EXL3 had42/48 substantive and6 refusals/evasions, including partial evasion. All14 matched cap-resolution answers reached EOS. These are dependent language/mode conditions and model-agent judgments, not48 independent tasks or independent human review. Substantive does not mean correct; language, factual and format defects remain. Remaining Chinese political evasions mean a broad uncensoring claim is unsupported. See behavioral details.

These results describe the unchanged v1 weights and packaged adapter0.1.3. Later experimental replacement probes and separately tested adapter0.1.4 are not part of this model revision. The audited abliteration operation is not a promise that every unnecessary refusal is removed. Original glue redistribution licensing is still unresolved; this update grants no new licence.