---
library_name: transformers
base_model: WaveCut/DeepSeek-V4-Flash-0731-REAM128-146B
license: mit
tags:
- deepseek-v4
- mixture-of-experts
- ream
- fp4
- fp8
---

<!-- ream-line-status -->
> [!WARNING]
> **Experimental — most likely broken (renamed to `-exp` on 2026-08-12).** This variant was produced by the second cascade step (REAM160 → REAM128 → REAM96). Severe multi-turn degradation was observed: generation typically collapses into gibberish from the second request onward. No quality evidence was ever collected for this release. Kept for archival reference only; do not use for inference. A rebuilt REAM96 (single-step prune from the base model, new calibration) will be published at the original repo id `WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B`.

# DeepSeek V4 Flash 0731 REAM96-111B

This release is a calibration-guided, training-free cascade from the committed
REAM128 checkpoint. It reduces every routed MoE layer from 128 to 96 experts
while preserving 43 transformer layers, one shared expert, and top-k 6 routing.
The logical parameter count is 111,163,180,631.

## Lineage and precision

- Source: `WaveCut/DeepSeek-V4-Flash-0731-REAM128-146B@84dc02df6c844568738da6bc425ea4772a7c56f4`.
- Source checkpoint identity: `f96d633ad1841c18f8f3cdd960cd64c7533aee46`.
- Source `NATIVE_READY.json` SHA-256: `8c32c76316574c75f18b45be69bd3b54c0c547531249b02f424a66485dc4820e`.
- Run: `20260811T072216Z`.
- Build code revision: `47b3fe27862537b0a9504a135d9728fba54e0596`.
- Container: `sha256:0a360022e8de4375af99430f84e8b38951acc397252163a37ceac7204d01be35`.
- Routed expert weights retain the source-native E2M1 FP4 layout with E8M0 scales.
- Attention projections and unchanged tensors retain the source FP8/mixed layout.
- REAM accumulation is deterministic FP32; BF16 is transient compute, not the primary checkpoint format.
- MTP is disabled (`num_nextn_predict_layers: 0`).
- Checkpoint: 63,982,995,880 bytes across 44 Safetensors shards, `config.json`, and `model.safetensors.index.json`.
- Committed `NATIVE_READY.json` SHA-256: `1a0aef0c8d4f736516ec1268b08d3ad698bef16f8e0ee90739fc0a0032831527`.

## REAM methodology

REAM96 does not apply gradient updates or add training examples to the
checkpoint. Fresh layerwise saliency statistics, expert groups, merges, router
mappings, and activations were computed against the immutable REAM128 source.
The REAM128 importance matrix was not reused.

Calibration composition: 3,072 samples x 512 tokens.

| Slice | Samples | Share | Pinned sources and focus |
| --- | ---: | ---: | --- |
| Code | 614 | 20.0% | `sahil2801/CodeAlpaca-20k` and `bigcode/self-oss-instruct-sc2-exec-filter-50k`; instruction following and executable code |
| Math | 307 | 10.0% | `openai/gsm8k` and `AI-MO/NuminaMath-CoT`; arithmetic and mathematical reasoning |
| Agent/tool | 1,075 | 35.0% | `NousResearch/hermes-function-calling-v1`; function selection and structured tool calls |
| Russian/Cyrillic | 1,076 | 35.0% | `ZeroAgency/ru-big-russian-dataset`; Russian and Cyrillic language coverage |

At each layer, REAP saliency is the routed activation norm multiplied by router
probability. The 96 highest-saliency source experts become centroids; the
remaining 32 source experts are assigned using output and router-gate
similarity. Each group is merged by saliency-weighted deterministic FP32
accumulation, router rows are remapped, and top-k 6 routing is preserved.

## Validation and integrity

- Build hardware: 2x NVIDIA GeForce RTX 4090 24 GB.
- Structural gate: 43 layers, 96 routed experts, one shared expert, top-k 6, and no MTP layers.
- Native release gate: all 46 checkpoint artifacts match the committed byte sizes and SHA-256 values.
- The native checkpoint became terminal-ready after 21 h 34 min of the run, including recoveries and durability checks.
- Comparative heldout PPL/NLL: not measured for this cascade.
- Semantic generation suite: not run for this cascade.
- Runtime benchmark: not run for this cascade.

Artifact and lineage checks prove structural integrity; they are not quality or
performance evidence.

## Loading and runtime compatibility

The configuration and tokenizer can be inspected from the public repository:

```python
from transformers import AutoConfig, AutoTokenizer

repo_id = "WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B"
config = AutoConfig.from_pretrained(repo_id, trust_remote_code=False)
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=False)
print(config.num_hidden_layers, config.n_routed_experts, config.num_experts_per_tok)
```

Stock `AutoModelForCausalLM`, vLLM, and SGLang generation is not claimed for
this packed E2M1/E8M0 checkpoint. A runtime must implement the 96-expert
DeepSeek V4 topology and its native FP4/FP8 storage.

## Limitations

- Quality and throughput must not be inferred from the REAM128 release; this cascade has no PPL, NLL, semantic, or benchmark result.
- This is a specialized mixed-precision checkpoint, not a conventional BF16/F16 Transformers checkpoint.
- Hardware memory requirements depend on the runtime, sharding strategy, context length, and KV-cache configuration.

## License

MIT License, following the source checkpoint.