DeepSeek V4 Flash
DeepSeek · deepseek-ai/DeepSeek-V4-Flash
Q4 file 137.9 GB (UD-IQ4_XS) · runs well from 512 GB · Weights license: MIT
Quality
Quality · Overall
LiveBench 2026-06-25 · 0–100 · higher is better · this model outlined, #21 of 22
- Color = publisher
- Memory under review
Source: LiveBench 2026-06-25 · collected 2026-09-26 · open models without a score are not drawn (no score ≠ 0)
- LMArena · text · rating
- 1,436 ± 4Confidence interval 1,432–1,440 · 51,982 votes · rank 105 of the releaseLMArena 2026-10-22 · CC BY 4.0
LiveBench categories
This model's mean per category · 0–100
Source: LiveBench 2026-06-25 · collected 2026-09-26 · open models without a score are not drawn (no score ≠ 0)
Reported by the publisher
- MMLU-Pro (EM)
- 83.0V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- SimpleQA-Verified (Pass@1)
- 23.1V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- Chinese-SimpleQA (Pass@1)
- 71.5V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- GPQA Diamond (Pass@1)
- 71.2V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- HLE (Pass@1)
- 8.1V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- LiveCodeBench (Pass@1)
- 55.2V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- HMMT 2026 Feb (Pass@1)
- 40.8V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- IMOAnswerBench (Pass@1)
- 41.9V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
Every reported result
- Apex (Pass@1)
- 1.0V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- Apex Shortlist (Pass@1)
- 9.3V4-Flash Non-Think · Knowledge & Reasoning · thinking mode
- MRCR 1M (MMR)
- 37.5V4-Flash Non-Think · Long Context · thinking mode
- CorpusQA 1M (ACC)
- 15.5V4-Flash Non-Think · Long Context · thinking mode
- Terminal Bench 2.0 (Acc)
- 49.1V4-Flash Non-Think · Agentic · thinking mode
- SWE Verified (Resolved)
- 73.7V4-Flash Non-Think · Agentic · thinking mode
- SWE Pro (Resolved)
- 49.1V4-Flash Non-Think · Agentic · thinking mode
- SWE Multilingual (Resolved)
- 69.7V4-Flash Non-Think · Agentic · thinking mode
- MCPAtlas (Pass@1)
- 64.0V4-Flash Non-Think · Agentic · thinking mode
- Toolathlon (Pass@1)
- 40.7V4-Flash Non-Think · Agentic · thinking mode
- MMLU-Pro (EM)
- 86.4V4-Flash High · Knowledge & Reasoning
- SimpleQA-Verified (Pass@1)
- 28.9V4-Flash High · Knowledge & Reasoning
- Chinese-SimpleQA (Pass@1)
- 73.2V4-Flash High · Knowledge & Reasoning
- GPQA Diamond (Pass@1)
- 87.4V4-Flash High · Knowledge & Reasoning
- HLE (Pass@1)
- 29.4V4-Flash High · Knowledge & Reasoning
- LiveCodeBench (Pass@1)
- 88.4V4-Flash High · Knowledge & Reasoning
- Codeforces (Rating)
- 2,816V4-Flash High · Knowledge & Reasoning
- HMMT 2026 Feb (Pass@1)
- 91.9V4-Flash High · Knowledge & Reasoning
- IMOAnswerBench (Pass@1)
- 85.1V4-Flash High · Knowledge & Reasoning
- Apex (Pass@1)
- 19.1V4-Flash High · Knowledge & Reasoning
- Apex Shortlist (Pass@1)
- 72.1V4-Flash High · Knowledge & Reasoning
- MRCR 1M (MMR)
- 76.9V4-Flash High · Long Context
- CorpusQA 1M (ACC)
- 59.3V4-Flash High · Long Context
- Terminal Bench 2.0 (Acc)
- 56.6V4-Flash High · Agentic
- SWE Verified (Resolved)
- 78.6V4-Flash High · Agentic
- SWE Pro (Resolved)
- 52.3V4-Flash High · Agentic
- SWE Multilingual (Resolved)
- 70.2V4-Flash High · Agentic
- BrowseComp (Pass@1)
- 53.5V4-Flash High · Agentic
- HLE w/ tools (Pass@1)
- 40.3V4-Flash High · Agentic
- MCPAtlas (Pass@1)
- 67.4V4-Flash High · Agentic
- Toolathlon (Pass@1)
- 43.5V4-Flash High · Agentic
- MMLU-Pro (EM)
- 86.2V4-Flash Max · Knowledge & Reasoning
- SimpleQA-Verified (Pass@1)
- 34.1V4-Flash Max · Knowledge & Reasoning
- Chinese-SimpleQA (Pass@1)
- 78.9V4-Flash Max · Knowledge & Reasoning
- GPQA Diamond (Pass@1)
- 88.1V4-Flash Max · Knowledge & Reasoning
- HLE (Pass@1)
- 34.8V4-Flash Max · Knowledge & Reasoning
- LiveCodeBench (Pass@1)
- 91.6V4-Flash Max · Knowledge & Reasoning
- Codeforces (Rating)
- 3,052V4-Flash Max · Knowledge & Reasoning
- HMMT 2026 Feb (Pass@1)
- 94.8V4-Flash Max · Knowledge & Reasoning
- IMOAnswerBench (Pass@1)
- 88.4V4-Flash Max · Knowledge & Reasoning
- Apex (Pass@1)
- 33.0V4-Flash Max · Knowledge & Reasoning
- Apex Shortlist (Pass@1)
- 85.7V4-Flash Max · Knowledge & Reasoning
- MRCR 1M (MMR)
- 78.7V4-Flash Max · Long Context
- CorpusQA 1M (ACC)
- 60.5V4-Flash Max · Long Context
- Terminal Bench 2.0 (Acc)
- 56.9V4-Flash Max · Agentic
- SWE Verified (Resolved)
- 79.0V4-Flash Max · Agentic
- SWE Pro (Resolved)
- 52.6V4-Flash Max · Agentic
- SWE Multilingual (Resolved)
- 73.3V4-Flash Max · Agentic
- BrowseComp (Pass@1)
- 73.2V4-Flash Max · Agentic
- HLE w/ tools (Pass@1)
- 45.1V4-Flash Max · Agentic
- MCPAtlas (Pass@1)
- 69.0V4-Flash Max · Agentic
- GDPval-AA (Elo)
- 1,395V4-Flash Max · Agentic
- Toolathlon (Pass@1)
- 47.8V4-Flash Max · Agentic
Across DeepSeek V4 sizes
| Benchmark (publisher settings) | Flash 284B | Pro 1.6T |
|---|---|---|
| MMLU-Pro (EM) | 83.0 | 82.9 |
| SimpleQA-Verified (Pass@1) | 23.1 | 45.0 |
| Chinese-SimpleQA (Pass@1) | 71.5 | 75.8 |
| GPQA Diamond (Pass@1) | 71.2 | 72.9 |
| HLE (Pass@1) | 8.1 | 7.7 |
| LiveCodeBench (Pass@1) | 55.2 | 56.8 |
| HMMT 2026 Feb (Pass@1) | 40.8 | 31.7 |
| IMOAnswerBench (Pass@1) | 41.9 | 35.3 |
| Apex (Pass@1) | 1.0 | 0.4 |
| Apex Shortlist (Pass@1) | 9.3 | 9.2 |
| MRCR 1M (MMR) | 37.5 | 44.7 |
| CorpusQA 1M (ACC) | 15.5 | 35.6 |
| Terminal Bench 2.0 (Acc) | 49.1 | 59.1 |
| SWE Verified (Resolved) | 73.7 | 73.6 |
| SWE Pro (Resolved) | 49.1 | 52.1 |
| SWE Multilingual (Resolved) | 69.7 | 69.8 |
| MCPAtlas (Pass@1) | 64.0 | 69.4 |
| Toolathlon (Pass@1) | 40.7 | 46.3 |
| MMLU-Pro (EM) | 86.4 | 87.1 |
| SimpleQA-Verified (Pass@1) | 28.9 | 46.2 |
| Chinese-SimpleQA (Pass@1) | 73.2 | 77.7 |
| GPQA Diamond (Pass@1) | 87.4 | 89.1 |
| HLE (Pass@1) | 29.4 | 34.5 |
| LiveCodeBench (Pass@1) | 88.4 | 89.8 |
| Codeforces (Rating) | 2,816 | 2,919 |
| HMMT 2026 Feb (Pass@1) | 91.9 | 94.0 |
| IMOAnswerBench (Pass@1) | 85.1 | 88.0 |
| Apex (Pass@1) | 19.1 | 27.4 |
| Apex Shortlist (Pass@1) | 72.1 | 85.5 |
| MRCR 1M (MMR) | 76.9 | 83.3 |
| CorpusQA 1M (ACC) | 59.3 | 56.5 |
| Terminal Bench 2.0 (Acc) | 56.6 | 63.3 |
| SWE Verified (Resolved) | 78.6 | 79.4 |
| SWE Pro (Resolved) | 52.3 | 54.4 |
| SWE Multilingual (Resolved) | 70.2 | 74.1 |
| BrowseComp (Pass@1) | 53.5 | 80.4 |
| HLE w/ tools (Pass@1) | 40.3 | 44.7 |
| MCPAtlas (Pass@1) | 67.4 | 74.2 |
| Toolathlon (Pass@1) | 43.5 | 49.0 |
| GDPval-AA (Elo) | 1,395 | 1,554 |
Values as printed on the card, with the publisher's own settings (card section: Evaluation Results › Instruct Model › Comparison across Modes). Not comparable across publishers and not used in any ranking. Model card
Run it
Run it
File UD-IQ4_XS · 137.9 GB · 4 parts · unsloth/DeepSeek-V4-Flash-GGUF · What is Q4?
Measured 39 tok/s · FP8 · 1 machine · Lambda inference models · all rows
llama-server -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-IQ4_XS- Install llama.cpp
- Run the command above
- Ask at localhost:8080:
Explain what a context window is, in two sentences.
Quality × memory
Quality × memory
LiveBench 2026-06-25 overall vs. Q4 planning memory (log scale) · 18 open models
- This model
- Color = publisher
Memory = the model's own figure: parameters × 0.625 bytes + 20% working room, before any machine's reserve; with a chosen machine, the total there with its reserve (25% of memory, 2 to 8 GB; 2 GB per graphics card). An estimate, not measured VRAM. LiveBench 2026-06-25
Runs on
Fast memory holds it 1
The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.
Within memory, outside the fast part 1
Within the machine's memory, outside its fast part: on a graphics card the layers that do not fit run on the CPU from PC RAM (the listed RAM, or 32 GB assumed); on a Mac, above the macOS default GPU share. Not a speed figure.
One memory tier short 9
The same machines with the next memory size up would hold it in fast memory.
Show all 9
DeepSeek V4 Flash is a text model from DeepSeek with 284B parameters, 13B of them active per token (mixture of experts). It was released on April 24, 2026 and reads up to 1,048,576 tokens of context. Its Q4 weights plus working room come to about 213 GB before the machine's reserve; the smallest reference machine whose fast memory holds that is Mac Studio · M5 Ultra · 36-core CPU / 80-core GPU · 512 GB unified (~221 GB with its reserve). We collected 16 downloadable GGUF and MLX files from 2 publisher sources; the most downloaded GGUF source is unsloth's (234.5K downloads on the Hub). It scores 65.5 overall in LiveBench 2026-06-25. Its LMArena text rating is 1,436 ± 4 (release 2026-10-22), a separate human-preference scale.
Hardware profiles grouped by the Q4 planning estimate on each machine (dedicated GPUs reserve 2 GB; unified and system memory 25%, 2 to 8 GB): the fast memory holds it, it is within memory but outside the fast part (layers offloaded to PC RAM, or above the macOS GPU share), or the machine is one memory tier short. Each card opens its hardware profile. Not a measured fit or speed.
On the reference machines
| Where Q4 sits | ||||
|---|---|---|---|---|
| Mac mini M2 · 8 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac mini M4 · 16 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~217 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac mini M4 · 24 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~219 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac mini M4 · 32 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~221 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac mini M5 Pro · 64 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~221 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac Studio M5 Max · 128 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~221 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| Mac Studio · M5 Ultra · 36-core CPU / 80-core GPU · 512 GB unifiedFast memory holds itUD-IQ4_XSKV-cache formula for this architecture not reviewed | ~221 GB | Fast memory holds it | UD-IQ4_XS | KV-cache formula for this architecture not reviewed |
| RTX 4060 · 8 GB VRAMLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| RTX 3060 · 12 GB VRAMLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| RTX 5080 · 16 GB VRAMLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| RTX 4090 · 24 GB VRAMLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| RTX 5090 · 32 GB VRAMLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column | ~215 GB | Larger than its memory | No file on the site's quant ladder fits the fast memory at 8,192 tokens | No file in the previous column |
| DGX H100 · 8 × 80 GBFast memory holds itUD-IQ4_XSKV-cache formula for this architecture not reviewed | ~229 GB | Fast memory holds it | UD-IQ4_XS | KV-cache formula for this architecture not reviewed |
The file column uses the site's quant ladder (Q8_0 to Q3_K_M, MLX 8 to 4-bit on a Mac): the largest collected file whose size × 1.2 plus the context cache and the machine's reserve fits the fast memory. Planning estimates, not measured runs.
Compare DeepSeek V4 Flash using its publisher card and the LiveBench 2026-06-25 snapshot. This is quality evidence, not local speed.
The Q4 weights plus working room come to about 213 GB from 284B stated parameters, before the machine's reserve (Runs on adds it per machine). It is not a measured minimum; context, batch size, runtime and quantization change actual memory use.
Versions to download
Files you can download for this checkpoint. The section opens on the most downloaded collected GGUF source, unsloth (234.5K Hub downloads), at UD-IQ4_XS. On a Mac it opens on the most downloaded MLX source. Sizes are the Hub's published bytes. What is Q4?
unsloth · 234.5K downloads
unsloth/DeepSeek-V4-Flash-GGUF · UD-IQ4_XS/DeepSeek-V4-Flash-UD-IQ4_XS-00001-of-00004.gguf · 137.9 GB
Run this file
llama-server -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-IQ4_XSlms get https://huggingface.co/unsloth/DeepSeek-V4-Flash-GGUF- Ollama: Ollama's documented hf.co form covers single-file GGUFs; this file has 4 parts
- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Quality scores on this page belong to the original checkpoint; these files are not scored or ranked. File sizes read 2026-10-23 · download counts read 2026-10-17 · commands: Hugging Face Hub docs: Use Ollama with any GGUF model · llama.cpp server README (-hf <user>/<model>[:quant]) · LM Studio docs: lms get with a Hugging Face URL · mlx-lm README (mlx_lm.generate --model)
Fine-tunes, adapters and other versions
- Quantized versions
- 134
- Fine-tunes
- 28
- Adapters · merges
- 4 · 0
Browse them on the Hugging Face model tree · read 2026-10-23
Uncensored and abliterated versions
Some community uploads remove the model's safety refusals. Quality may drop, the original publisher does not support them, claims about them are the uploader's, and the original license still applies. This site never recommends or ranks them.
15 derivative repositories have “uncensored” or “abliterated” in their name (read 2026-10-23).
Measured speed
Tokens per second by machine
Output tok/s · 1 machine · every row in the table · What is a token?
| DGX H100 | 39 per user (32 concurrent) | FP8 (sgl-project/DeepSeek-V4-Flash-FP8) | SGLang with EAGLE speculative decoding (docker lmsysorg/sglang:deepseek-v4-hopper) | sgl-project/DeepSeek-V4-Flash-FP8 (FP8 repack, about 284 GB); not deepseek-ai/DeepSeek-V4-Flash |
|---|
Source: Lambda inference models · DeepSeek-V4-Flash (vendor benchmark) (checked 2026-10-09) · SGLang with EAGLE speculative decoding (docker lmsysorg/sglang:deepseek-v4-hopper) · docker lmsysorg/sglang:deepseek-v4-hopper (no SGLang version number); SGLANG_DSV4_FP4_EXPERTS=0; --tp 8; --mem-fraction-static 0.85; --speculative-algo EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4; --chunked-prefill-size 4096 · not estimated
How to run
The memory needed above is the Q4 load before a machine's reserve. Formula: parameters × 0.625 bytes, plus 20% working allowance; each machine then adds its system reserve: 25% of unified or system memory, 2 to 8 GB, or 2 GB per GPU on dedicated cards. This is neither measured VRAM nor checkpoint download size.
Q4 planning estimate only. Check checkpoint format, runtime support, KV cache, context, batch size and offload before choosing hardware.
Evidence level: Generic estimate only; no execution reproduced by this site.
- Publisher sampling defaults
- temperature 1 · top_p 1generation_config.json · read 2026-10-23source
- Chat template markers
- — Chat template not read for this checkpoint
Before downloading
- Check the exact checkpoint and license linked here.
- Confirm a compatible runtime and precision or quantization for your platform.
- Allow for context, batch size and any extra media components beyond the weights.
Technical details
What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.
- Checkpoint / version
- deepseek-ai/DeepSeek-V4-Flash
- Task and modality
- language and reasoning · text
- Count and basis
- 284B · manufacturer specification · 13B active
- Context
- 1,048,576 tokens published for this model
- Architecture
- MoE
- Released
- Change Log | DeepSeek API Docs · read 2026-10-09source
- Runtime and precision
- No local execution verified.
- Commercial use
- Allowed MIT license text (SPDX) · read 2026-10-18license text
Every stored field
- Revision checked
- 60d8d70770c6776ff598c94bb586a859a38244f1
- Architecture
- DeepseekV4ForCausalLM
- Library
- transformers
- Hub task tag
- text-generation
- Formats
- safetensors
- Access
- Open download
- Publisher kind
- Official
- Repository created
- 2026-04-22creation is not the model's release
- Repository updated
- 2026-06-22a commit is not a release date
- Hub downloads (30 days) · likes
- 936,089 · 2,344this repository only · read 2026-10-25
- Catalog state
- Hub config or weights found; no inventory benchmark link
Sources
Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.
- Publisher checkpoint and model card
- Weight license and use terms
- Quality benchmark source · LiveBench 2026-06-25 overall
Catalog reviewed: 2026-09-30. Benchmark release: 2026-06-25; snapshot collected: 2026-09-26.








