[!TIP] Support this work → · X · GitHub · REAP paper · Cerebras REAP

DeepSeek-V4-Flash-162B

REAP-pruned deepseek-ai/DeepSeek-V4-Flash.

At a glance

Which variant should I pick?

162B parameters | K144 REAP-pruned | 200K context | no speculative decoding

This is a smaller pruned DeepSeek V4 Flash that runs on a single DGX Spark. It trades some model capacity for higher prefill speed and a more conservative memory profile. It is the fallback option when you want 200K context with a bit more headroom.

What this is

• Base: deepseek-ai/DeepSeek-V4-Flash
• Pruning: REAP (Routing-Enhanced Activation Pruning) at K144
• Final size: ~162B active parameters
• Quantization: NVFP4 / MXFP4 expert weights with FP8 KV cache
• Serving: vLLM with DeepSeek V4 tokenizer, reasoning parser, and tool-call parser
• Context: 200,000 tokens validated end-to-end
• Hardware target: single NVIDIA DGX Spark / GB10 / SM121

K144 was the smaller checkpoint that still reached 200K on one Spark. It prefills faster than K160 (about 539 tok/s vs 514 tok/s) but decodes slower (about 14 tok/s vs 24 tok/s) because it lacks MTP speculative decoding. The watchdog also logged a low-memory kill at final teardown, so treat this as proof-of-concept rather than a comfortable always-on daemon.

How the REAP checkpoint was made

REAP (Router-weighted Expert Activation Pruning) is the Cerebras Research one-shot MoE compression method: https://github.com/CerebrasResearch/reap.

Short version: take DeepSeek V4 Flash, measure which MoE experts actually matter under real prompts, keep the most useful routed experts, delete the colder ones, remap the router/expert tables, then pack the surviving model into the low-bit format we serve.

Step by step:

1. Start from DeepSeek V4 Flash. DeepSeek V4 Flash is a sparse MoE model. Every token does not use every expert; the router picks a small top-k subset per token. That sparsity is what makes expert pruning viable. The served K144 checkpoint keeps this structure: model_type=deepseek_v4, 43 hidden layers, hidden size 4096, 1 shared expert, 6 routed experts active per token, and max_position_embeddings=1048576 from the base.

2. Run calibration prompts through the original model. A calibration corpus is passed through the unpruned model. For each token and each MoE layer, REAP records router scores, which experts the top-k selected, how strongly the router weighted them, and how large the expert activations were. The useful signal is roughly router_probability * topk_selected * activation_strength * frequency. This is the "router-weighted activation" part of the name.

3. Rank experts per layer. Each MoE layer gets its own ranking. Hot experts are ones the router actually depends on; cold experts are rarely picked or contribute little.

for layer in moe_layers:
    scores = {}
    for batch in calibration_data:
        router_output = model.router(layer, batch.hidden_states)
        topk_experts, gate_weights = select_experts(router_output)
        for token in batch.tokens:
            for expert, weight in topk_experts[token]:
                activation = estimate_activation_strength(layer, expert, token)
                scores[expert] += weight * activation
    keep_experts[layer] = top_k(scores, K)

For this checkpoint, K=144 routed experts per MoE layer are kept. The shared expert is always kept.

4. Physically prune the expert weights. This is structural surgery on the MoE expert tensors, not LoRA, prompt tuning, or fine-tuning. Embeddings, attention, norms, router, shared expert, selected routed experts, and the LM head all stay. Low-ranked routed experts are removed and expert IDs are remapped so the model has a compact expert table. That is why the config now reports n_routed_experts: 144 instead of the larger original count.

5. Update router metadata. Because experts were deleted, the router cannot point at old expert IDs. REAP rewrites the routing metadata and the token-to-expert mapping used by the runtime. This is why vLLM needed a router patch: K144 and K160 are valid checkpoints but use nonstandard routed-expert counts that some fused CUDA router kernels do not template-instantiate. The patch forces the general fallback router path. It does not change weights or model behavior.

6. Quantize and pack. The pruned checkpoint is packed into the low-bit format the runtime serves: MXFP4/NVFP4-style packed expert weights with FP8 MLA KV cache. That is how K144 lands in a memory range that fits on one DGX Spark with extra prefill headroom.

7. Validate quality and fit. Multiple sizes were tested on one Spark: 213B was too large, 200B failed readiness, 180B/K160 was the best balance, and 162B/K144 is the smaller fallback profile published here.

What REAP changes vs. preserves

Changes: number of routed experts, expert tensors, expert ID mapping, checkpoint size, runtime memory footprint.

Preserves: context length, tokenizer, attention architecture, number of layers, hidden size, number of experts used per token, base chat format.

What we did in this project

We did not recreate the REAP pipeline ourselves. We downloaded the already-created REAP checkpoints, inspected their configs and expert counts, patched vLLM to accept the nonstandard expert counts, built and validated the DGX Spark runtime, found working one-Spark profiles, and published the serving repos, configs, and model cards.

The end-to-end artifact:

DeepSeek V4 Flash
  -> router-weighted expert pruning (REAP)
  -> K144 expert-retained checkpoint
  -> low-bit packed checkpoint
  -> vLLM Spark runtime
  -> 200K context serving recipe

How we got here

See the DeepSeek-V4-Flash-Spark model card for the full story. The short version: we tested every REAP checkpoint from 148B through 213B on a single DGX Spark. Most failed before the API came up. K160 was the largest that survived with speculative decoding. K144 was the next viable option without it.

The same runtime patches apply: native ARM64 vLLM build, Cutlass 4.5.1 workaround, REAP expert-count fallback, MXFP4 memory hygiene, and FlashInfer CUDA IPC fix.

One-command install

Run this on the DGX Spark. HF_TOKEN is only needed if the model repo is private or not already cached.

HF_TOKEN=... bash -lc 'set -euo pipefail; cd /home/sero/spark; rm -rf deepseek-spark; git clone https://github.com/0xSero/deepseek-spark.git; cd deepseek-spark; ./setup.sh full k144'

Do not commit tokens. Pass them only through the environment for this one command.

Exact working profile

The profile lives at configs/k144-nospec-200k.env in the GitHub repo.

MODEL_REPO=0xSero/DeepSeek-V4-Flash-162B
MODEL_REVISION=d663e8fb16809f6619000648b187b257249ed824
SERVED_MODEL_NAME=DeepSeek-V4-Flash-Spark-Mini
CONTEXT_LENGTH=200000
KV_CACHE_MEMORY_BYTES=14G
MAX_NUM_BATCHED_TOKENS=8192
MAX_NUM_SEQS=1
GPU_MEMORY_UTILIZATION=0.88
WATCHDOG_MIN_AVAILABLE_KB=8388608
KV_CACHE_DTYPE=fp8
THINKING=true
SPECULATIVE_CONFIG=
VLLM_ENABLE_DEEPSEEK_V4_SPARSE_MLA_WARMUP=0
VLLM_TRITON_MLA_SPARSE_ALLOW_CUDAGRAPH=1

The launcher enables DeepSeek V4 tokenizer, reasoning parser, tool-call parser, prefix caching, FP8 KV, and CUDA graph capture. No speculative decoding.

Docker runtime

The runtime Docker image is published at:

ghcr.io/0xsero/deepseek-v4-flash-spark-vllm:cutlass451-g27

The image lineage is the DGX Spark DeepSeek V4 vLLM build vllm-node-dsv4:latest with vLLM 0.1.dev17016+g27fd665bd.d20260526 and nvidia-cutlass-dsl[cu13]==4.5.1. The final local tag is vllm-node-dsv4-cutlass451:latest.

Exact image validated on spark-2822:

vllm-node-dsv4-cutlass451:latest
sha256:5df60ebb9c10dfb86d5946cae8244adfe65a7fd405401bd542ecf22d5c497a4a

The installer pulls the published image automatically. Pass IMAGE_REF=... only when testing a different runtime image.

The runtime patcher applies the nonstandard REAP expert-count router fallback, MXFP4 memory hygiene, optional cute-dsl override hook, and a FlashInfer CUDA IPC libcudart fix. It does not modify model weights.

Validation

Run on spark-2822, a single DGX Spark / GB10 / SM121, on May 27 2026.

200K long-needle benchmark:

run_dir: /home/sero/spark/benchmarks/deepseek-reap/single-server-sweep/k144-nospec-200k-mnbt8192-20260527T190139Z
prompt_tokens: 186,390
TTFT: 345.834 s
prefill: 538.958 tok/s
decode: 13.899 tok/s
needle_retained: true

Task coverage at 200K included smoke, ASCII, Unicode, and Mermaid diagrams; code explanation; religion and philosophy prompts; tool-call fidelity; and long-needle retrieval. All passed. The watchdog logged a low-memory kill at final teardown near the 8 GB threshold, so this is proven but not the most comfortable always-on profile.

K144 with MTP2 was tested but was not long-context safe at the tested watchdog thresholds. The published 200K profile is therefore the no-speculative-decoding profile.

Why K144 without speculative decoding

K144 without MTP is the conservative option. It uses a larger KV cache (14 GB vs 6 GB) and bigger prefill chunks (8192 vs 4096), which gives it the highest prefill speed of the tested single-Spark profiles. The tradeoff is lower decode speed and a tighter memory margin at teardown.

Choose this if you value prefill throughput over decode speed, or if you want a simpler profile without speculative decoding.

Limitations

• This is a pruned model. It is not the full DeepSeek V4 Flash. Evaluate quality against your own tasks before trusting it for production work.
• 200K context works, but memory is tight. The watchdog killed the process at teardown during validation.
• The public 200K success path for the full model remains dual-Spark TP=2. This is a compromise.
• The Docker image and patches are experimental. They are not upstream vLLM and may break on newer commits.

Links

• One-command wrapper: https://github.com/0xSero/deepseek-spark
• Runtime module (configs, patcher, evidence): https://github.com/0xSero/deepseek-spark/tree/main/runtime
• Base model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
• Larger single-Spark profile: https://huggingface.co/0xSero/DeepSeek-V4-Flash-180B

License

MIT for the serving recipe and tooling. The base model weights follow the DeepSeek V4 Flash license. Review it before use.

License & citation

License inherited from the base model.

@misc{lasby2025reap,
  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
  year   = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}

Sponsors

Made possible by NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle.