# DeepSeek-V4-Flash-DSpark by DeepSeek: Open-Weight Model Source: https://savrn.com/models/deepseek-v4-flash-dspark Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve DeepSeek-V4-Flash-DSpark (165.3B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 330.5 GB | 396.6 GB | 2x [MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) (256 GB) Vultr | $4.00 | [2x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $5.18 · [3x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $5.55 | | 8-bit | 165.3 GB | 198.3 GB | 1x [MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) (256 GB) Vultr | $2.00 | [1x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $2.59 · [2x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $3.70 | | 4-bit | 82.6 GB | 99.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 · [1x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $2.59 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 24, 2026. [DeepSeek-V4-Flash-DSpark on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/deepseek-v4-flash-dspark/gpus) ## SAVRN's Notes on DeepSeek-V4-Flash-DSpark The publisher says it outright: DSpark is not a new model. It is the DeepSeek-V4-Flash checkpoint with a speculative decoding module attached, so you are evaluating a serving path, not fresh weights. Those weights run 165.3 billion parameters across 256 routed experts with 6 active per token. Precision picks the hardware: 16-bit needs 396.6 GB and two MI325X cards at $4.00 per hour, 8-bit needs 198.3 GB on one MI325X at $2.00, and 4-bit needs 99.2 GB on one MI300X at $1.85. MIT covers it: commercial use, modification and redistribution with the notices kept. Two checks before buying cards. The memory figures are the floor; a request that uses the 1,048,576-token context adds cache on top, sized from the single key/value head at 512 dimensions. And the speculative decoding path is the point of this release, so confirm your serving stack runs the publisher's inference example. SAVRN Research, 2026-09-18 ## Model Card By DeepSeek, published under mit, revision 62af8fffb2f7. ### DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence [Technical Report](https://savrn.com/papers/deepseek-v4-towards-highly-efficient-million-token-context-intelligence) Note: DeepSeek-V4-Flash-DSpark is not a new model. It is the same checkpoint with an additional speculative decoding module attached. A minimal inference example is available in the inference folder. For more details, refer to: https://github.com/deepseek-ai/DeepSpec ### Introduction We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: [Read the full model card (2,036 words)](https://savrn.com/models/deepseek-v4-flash-dspark/card) ## Configuration Architecture DeepseekV4ForCausalLM Context length (tokens) 1,048,576 Layers 43 Hidden size 4,096 Attention heads 64 Key/value heads 1 Head dimension 512 Vocabulary size 129,280 Routed experts 256 Experts active per token 6 Sliding window (tokens) 128 RoPE base 10,000 Stored precision bfloat16 Model type deepseek_v4 Quantization fp8 ## Identity and Version Repository deepseek-ai/DeepSeek-V4-Flash-DSpark Publisher DeepSeek Task Text generation Modality Text Library transformers Parameters 165.3B parameters Languages Not stated by the source Revision 62af8fffb2f7030cac4de2f0169f5b8d1101b646 First published 2026-06-27 Last updated 2026-07-04 ## Files and Weights 74 files, 166.9 GB in total. The weights are 48 files totalling 166.9 GB in safetensors. Weights48 files · 166.9 GB Configuration14 files · 5.7 MB Tokenizer2 files · 6.4 MB Documentation4 files · 24.5 KB Other5 files · 8.7 KB Repository1 file · 1.5 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model-00001-of-00048.safetensors | Weights | 1.1 GB | 517658661390 | | model-00002-of-00048.safetensors | Weights | 3.6 GB | 0821952bc12c | | model-00003-of-00048.safetensors | Weights | 3.6 GB | 2bd2b020f81f | | model-00004-of-00048.safetensors | Weights | 3.6 GB | e6286ec182a2 | | model-00005-of-00048.safetensors | Weights | 3.6 GB | 021dd40791d0 | | model-00006-of-00048.safetensors | Weights | 3.6 GB | 3d96bfc9686d | | model-00007-of-00048.safetensors | Weights | 3.6 GB | 4481b0c86e9f | | model-00008-of-00048.safetensors | Weights | 3.6 GB | e19559e75d35 | | model-00009-of-00048.safetensors | Weights | 3.6 GB | ab44718d102e | | model-00010-of-00048.safetensors | Weights | 3.6 GB | 743c16a35f1c | | model-00011-of-00048.safetensors | Weights | 3.6 GB | 4fc2c700363b | | model-00012-of-00048.safetensors | Weights | 3.6 GB | 5214de3c6ca8 | | model-00013-of-00048.safetensors | Weights | 3.6 GB | d953f118488b | | model-00014-of-00048.safetensors | Weights | 3.6 GB | 40c97b77151b | | model-00015-of-00048.safetensors | Weights | 3.6 GB | 136cfaa14f36 | | model-00016-of-00048.safetensors | Weights | 3.6 GB | 6881f71ac5da | | model-00017-of-00048.safetensors | Weights | 3.6 GB | a1949b35de69 | | model-00018-of-00048.safetensors | Weights | 3.6 GB | 8288d4462015 | | model-00019-of-00048.safetensors | Weights | 3.6 GB | 25a4d2a2a4a0 | | model-00020-of-00048.safetensors | Weights | 3.6 GB | 0016a46d9f85 | | model-00021-of-00048.safetensors | Weights | 3.6 GB | 3f342f4120b9 | | model-00022-of-00048.safetensors | Weights | 3.6 GB | 0bf9870e1d44 | | model-00023-of-00048.safetensors | Weights | 3.6 GB | 215f51b475b8 | | model-00024-of-00048.safetensors | Weights | 3.6 GB | a11f085685e7 | | model-00025-of-00048.safetensors | Weights | 3.6 GB | 1d2a9ec82956 | | model-00026-of-00048.safetensors | Weights | 3.6 GB | aaec7eb32b16 | | model-00027-of-00048.safetensors | Weights | 3.6 GB | 63069bb430af | | model-00028-of-00048.safetensors | Weights | 3.6 GB | 2a804253c04f | | model-00029-of-00048.safetensors | Weights | 3.6 GB | 68dfa669a272 | | model-00030-of-00048.safetensors | Weights | 3.6 GB | dd8433e27dff | | model-00031-of-00048.safetensors | Weights | 3.6 GB | 6e7796ad3b05 | | model-00032-of-00048.safetensors | Weights | 3.6 GB | 86d3f021bec5 | | model-00033-of-00048.safetensors | Weights | 3.6 GB | 35a92fa5a3cc | | model-00034-of-00048.safetensors | Weights | 3.6 GB | 368896471e8e | | model-00035-of-00048.safetensors | Weights | 3.6 GB | 85c9a9af1ec5 | | model-00036-of-00048.safetensors | Weights | 3.6 GB | 0e22e7f41331 | | model-00037-of-00048.safetensors | Weights | 3.6 GB | 47f41e12ce98 | | model-00038-of-00048.safetensors | Weights | 3.6 GB | cdbb8c36da34 | | model-00039-of-00048.safetensors | Weights | 3.6 GB | 03f3cc8ab1a7 | | model-00040-of-00048.safetensors | Weights | 3.6 GB | 3124105d3834 | | model-00041-of-00048.safetensors | Weights | 3.6 GB | 4eda737587c4 | | model-00042-of-00048.safetensors | Weights | 3.6 GB | 9ac23e2e32c4 | | model-00043-of-00048.safetensors | Weights | 3.6 GB | 8c629a864a9b | | model-00044-of-00048.safetensors | Weights | 3.6 GB | 745476d81cb1 | | model-00045-of-00048.safetensors | Weights | 1.1 GB | 9a0fd242134e | | model-00046-of-00048.safetensors | Weights | 3.6 GB | 14810f274692 | | model-00047-of-00048.safetensors | Weights | 3.6 GB | 7a44164698d9 | | model-00048-of-00048.safetensors | Weights | 3.7 GB | a0bbb24f36d2 | | config.json | Configuration | 1.9 KB | — | | encoding/encoding_dsv4.py | Configuration | 27.9 KB | — | | encoding/test_encoding_dsv4.py | Configuration | 3.7 KB | — | | encoding/tests/test_input_1.json | Configuration | 2.7 KB | — | | encoding/tests/test_input_2.json | Configuration | 526 B | — | | encoding/tests/test_input_3.json | Configuration | 4.5 KB | — | | encoding/tests/test_input_4.json | Configuration | 2.7 KB | — | | generation_config.json | Configuration | 170 B | — | | inference/config.json | Configuration | 1.2 KB | — | | inference/convert.py | Configuration | 6.6 KB | — | | inference/generate.py | Configuration | 5.8 KB | — | | inference/kernel.py | Configuration | 22.2 KB | — | | inference/model.py | Configuration | 45.1 KB | — | | model.safetensors.index.json | Configuration | 5.6 MB | — | | LICENSE | Documentation | 1.1 KB | — | | README.md | Documentation | 14.4 KB | — | | encoding/README.md | Documentation | 8.1 KB | — | | inference/README.md | Documentation | 951 B | — | | encoding/tests/test_output_1.txt | Other | 2.4 KB | — | | encoding/tests/test_output_2.txt | Other | 342 B | — | | encoding/tests/test_output_3.txt | Other | 3.3 KB | — | | encoding/tests/test_output_4.txt | Other | 2.6 KB | — | | inference/requirements.txt | Other | 92 B | — | | .gitattributes | Repository | 1.5 KB | — | | tokenizer.json | Tokenizer | 6.4 MB | — | | tokenizer_config.json | Tokenizer | 801 B | — | ## License and Download License mit Access Open weights, no gate Download size 166.9 GB [Download from DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark) Released by DeepSeek through its official repository on Hugging Face. [Read the license](https://opensource.org/license/mit). ## Built From - Described by [arXiv:2606.19348](https://savrn.com/papers/deepseek-v4-towards-highly-efficient-million-token-context-intelligence) ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 166.9 GB | | 16-bit | 330.5 GB | | 8-bit | 165.3 GB | | 4-bit | 82.6 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Compare DeepSeek-V4-Flash-DSpark - [DeepSeek-V4-Flash-DSpark vs gpt-oss-120b](https://savrn.com/models/compare/deepseek-v4-flash-dspark-vs-gpt-oss-120b) - [DeepSeek-V4-Flash-0731 vs DeepSeek-V4-Flash-DSpark](https://savrn.com/models/compare/deepseek-v4-flash-0731-vs-deepseek-v4-flash-dspark) ## Questions About DeepSeek-V4-Flash-DSpark ### How much GPU memory does DeepSeek-V4-Flash-DSpark need? About 396.6 GB at 16-bit and 99.2 GB at 4-bit: the weights (165.3B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run DeepSeek-V4-Flash-DSpark on? At 16-bit, 2x MI325X from $4.00 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### Can I use DeepSeek-V4-Flash-DSpark commercially? Yes. DeepSeek-V4-Flash-DSpark is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included. ### What is DeepSeek-V4-Flash-DSpark's context length? 1,048,576 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Text generation ### [MiMo-V2.6-Flash-RL](https://savrn.com/models/mimo-v2-6-flash-rl) [Xiaomi MiMo](https://savrn.com/model-publishers/xiaomimimo) Scaling Reinforcement Learning Toward Self-Improvement MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of the MiMo-V2.6 series. The series is built to scale reinforcement learning toward self-improvement — scaling RL compute, environment diversity, and grader compute together, so the model keeps expanding its capability frontier through exploration and feedback. Key features include: - Native Omnimodal + Long Horizon: Text, image, video, and audio in one model; 1M tokens for long repositories, tool traces, and multi-session agent runs. - Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2): After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned… Open weights mit 159.4B parameters 1,048,576 tokens transformers [View model](https://savrn.com/models/mimo-v2-6-flash-rl) Model · Text generation ### [MiMo-V2.6-Flash-RL](https://savrn.com/models/mimo-v2-6-flash-rl-2) [Zachary Howard](https://savrn.com/model-publishers/servantofares) Scaling Reinforcement Learning Toward Self-Improvement MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of the MiMo-V2.6 series. The series is built to scale reinforcement learning toward self-improvement — scaling RL compute, environment diversity, and grader compute together, so the model keeps expanding its capability frontier through exploration and feedback. Key features include: - Native Omnimodal + Long Horizon: Text, image, video, and audio in one model; 1M tokens for long repositories, tool traces, and multi-session agent runs. - Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2): After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned… Open weights mit 159.4B parameters 1,048,576 tokens transformers [View model](https://savrn.com/models/mimo-v2-6-flash-rl-2) Model · Text generation ### [Hy3-Razor-154B-A18B-E96of192](https://savrn.com/models/hy3-razor-154b-a18b-e96of192) [Mingyang Song](https://savrn.com/model-publishers/nickyang) Hy3 with half of its routed experts removed by RAZOR, a training-free expert pruning method. Every MoE layer keeps 96 of its original 192 routed experts. No gradient updates or recovery training were applied: the retained weights are the base model's own weights. Pruning touches only the routed expert pool. Attention, shared experts, the embedding and the LM head are untouched, so the compute per token drops only by the share of expert FLOPs that the removed experts would have contributed. The other budget is Requires a Transformers build containing the native hyv3 implementation. Weights are bfloat16. RAZOR asks whether the surviving computation can replace an expert's function, rather… Open weights apache-2.0 153.8B parameters 262,144 tokens transformers [View model](https://savrn.com/models/hy3-razor-154b-a18b-e96of192) Model · Text generation ### [Qwen3.8-Flash-Next-oQ4e-mix6-mtp](https://savrn.com/models/qwen3-8-flash-next-oq4e-mix6-mtp) [Umixer](https://savrn.com/model-publishers/umixer) Goal: oQ4e speed and memory footprint, oQ5e-or-better quality reaching toward oQ6e. Runs at ~73–75 GB resident with ~82→66 tok/s decode on a 128 GB Apple Silicon host — the speed/RAM economics of a 4-bit build, at an accuracy tier equal to or above uniform oQ5e. How: a mixed-precision MLX build of Qwen3.8-Flash-Next, assembled by component-wise byte-copy from Jundot's native oQ quantizations — zero re-quantization in the entire package. Every tensor is byte-identical to its source, so quality is traceable, not re-derived. Uniform-bit quantization pays accuracy where it isn't needed (middle layers) and loses it where it hurts (first/last layers, n-gram table, attention). A Data-Free… Open weights other 180B parameters 262,144 tokens mlx [View model](https://savrn.com/models/qwen3-8-flash-next-oq4e-mix6-mtp) Model · Text generation ### [keys-GLM-5.3-EXL3-2.75BPW](https://savrn.com/models/keys-glm-5-3-exl3-2-75bpw) [Keys](https://savrn.com/model-publishers/drowzeys) Full GLM-5.3 (753B, glmmoedsa) quantized to EXL3 with a per-expert mixed bit width averaging 2.75 bpw on the routed experts, sized to serve on four NVIDIA DGX Sparks (GB10, 128 GB unified memory each) with room for a 200K-token KV pool (or 1M tokens with decode-context-parallel 4). Unmodified weights of zai-org/GLM-5.3 (no abliteration, no fine-tuning), quantized from the BF16 checkpoint. For the earlier uniform-width quant see The fastest way to serve this checkpoint is TensorFold's native glmmoedsa engine — tensor parallel 4, one rank per Spark, no vLLM in the serving path; a drafted reply equals a serial one. Recipe, one-shot launcher, benches and an AGENTS.md needs no extra draft model… Open weights other 138.1B parameters 1,048,576 tokens exllamav3 [View model](https://savrn.com/models/keys-glm-5-3-exl3-2-75bpw) Model · Text generation ### [Qwen3.8-Flash-Next-P48NVFP4-MoESQ](https://savrn.com/models/qwen3-8-flash-next-p48nvfp4-moesq) [IST Austria Distributed Algorithms and Systems Lab](https://savrn.com/model-publishers/ista-daslab) A W4A4 + paired-4:8 sparse compressed checkpoint of Qwen/Qwen3.8-Flash-Next, produced with MoESQ. The routed MoE expert weights are NVFP4 with paired-4:8 structured sparsity and are stored sparse: only the kept values plus a small mask are on disk. The target is NVIDIA Blackwell (SM100) sparse tensor cores. expert, 48 layers: 36 linear-attention and 12 sparse-attention, hyper-connections, per-layer n-gram embedding) weight including the sparsity mask and scales (2 bits per weight for the kept values). TP2). The per-layer n-gram embedding tables (95.4 GiB, BF16, unchanged) are lookup tables served from CPU memory, as in the base model. paired48nvfp4 MoE backend Links This checkpoint does not… Open weights other 134.7B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen3-8-flash-next-p48nvfp4-moesq) ## DeepSeek [All models and datasets](https://savrn.com/model-publishers/deepseek-ai) ## Versions - [62af8fffb2f7](https://savrn.com/models/deepseek-v4-flash-dspark/versions/62af8fffb2f7) · current 2026-09-18 ## Explore More - [All text generation models](https://savrn.com/models/tasks/text-generation) - [All models under mit](https://savrn.com/models/licenses/mit) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-19. - [DeepSeek website](https://www.deepseek.com/) - [Hugging Face record](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark) - [How the hub is built](https://savrn.com/model-hub/methodology)