WaveCut

DeepSeek-V4-Flash-0731-REAM128-146B-exp

Lineage and precision

Source: WaveCut/DeepSeek-V4-Flash-0731-REAM160-180B@7b7e03914a2cf1d676161ae18bf0f24c54adb931.
Run: 20260810T115212Z.
Build code revision: ec9da81a9bb12979973d834ffbe7dbf0a6292bad.
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: 82,392,019,049 bytes across 44 Safetensors shards, config.json, and model.safetensors.index.json.
Committed NATIVE_READY.json SHA-256: 8c32c76316574c75f18b45be69bd3b54c0c547531249b02f424a66485dc4820e.

REAM methodology

REAM128 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 REAM160 source.

Calibration composition: 3,072 samples x 512 tokens.

Table with columns: Slice, Samples, Share, Pinned sources and focus
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%

At each layer, REAP saliency is the routed activation norm multiplied by router probability. The 128 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: 4x NVIDIA GeForce RTX 4090 24 GB.
Structural gate: 43 layers, 128 routed experts, one shared expert, and top-k 6.
Native release gate: all 46 checkpoint artifacts match the committed byte sizes and SHA-256 values.
Comparative heldout PPL/NLL: not measured for this cascade.
Semantic generation suite: not run for this cascade.
Runtime benchmark: not run for this cascade.

The integrity checks prove lineage and artifact identity; they are not a claim of semantic equivalence to REAM160.

Loading and runtime compatibility

The configuration and tokenizer can be inspected with Transformers:

python
from transformers import AutoConfig, AutoTokenizer

repo_id = "WaveCut/DeepSeek-V4-Flash-0731-REAM128-146B"
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 128-expert DeepSeek V4 topology and its native FP4/FP8 storage.

Limitations

Quality and throughput must not be inferred from the REAM160 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.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

WaveCut

Model Tree

Base

WaveCut/DeepSeek-V4-Flash-0731-REAM160-180B

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Lineage and precision

Source: WaveCut/DeepSeek-V4-Flash-0731-REAM160-180B@7b7e03914a2cf1d676161ae18bf0f24c54adb931.
Run: 20260810T115212Z.
Build code revision: ec9da81a9bb12979973d834ffbe7dbf0a6292bad.
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: 82,392,019,049 bytes across 44 Safetensors shards, config.json, and model.safetensors.index.json.
Committed NATIVE_READY.json SHA-256: 8c32c76316574c75f18b45be69bd3b54c0c547531249b02f424a66485dc4820e.

REAM methodology

Calibration composition: 3,072 samples x 512 tokens.

Table with columns: Slice, Samples, Share, Pinned sources and focus
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%

Validation and integrity

Build hardware: 4x NVIDIA GeForce RTX 4090 24 GB.
Structural gate: 43 layers, 128 routed experts, one shared expert, and top-k 6.
Native release gate: all 46 checkpoint artifacts match the committed byte sizes and SHA-256 values.
Comparative heldout PPL/NLL: not measured for this cascade.
Semantic generation suite: not run for this cascade.
Runtime benchmark: not run for this cascade.

The integrity checks prove lineage and artifact identity; they are not a claim of semantic equivalence to REAM160.

Loading and runtime compatibility

The configuration and tokenizer can be inspected with Transformers:

python
from transformers import AutoConfig, AutoTokenizer

repo_id = "WaveCut/DeepSeek-V4-Flash-0731-REAM128-146B"
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)

Limitations

Quality and throughput must not be inferred from the REAM160 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.

DeepSeek-V4-Flash-0731-REAM128-146B-exp

README

Lineage and precision

REAM methodology

Validation and integrity

Loading and runtime compatibility

Limitations

License

Explore FriendliAI today

README

Lineage and precision

REAM methodology

Validation and integrity

Loading and runtime compatibility

Limitations

License