DeepSeek V4 Flash on Run BiOS

DeepSeek-V4-Flash is the fast, efficient model of the DeepSeek-V4 series, a family of Mixture-of-Experts (MoE) language models built by DeepSeek (DeepSeek-AI) and described in the technical report "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence" (2026). DeepSeek-V4-Flash has 284B total parameters with 13B activated per token, and supports a context length of one million tokens.

The checkpoint served on Run BiOS is DeepSeek-V4-Flash-0731, the official release of DeepSeek-V4-Flash published by DeepSeek on July 31, 2026. Per DeepSeek's official model card, DeepSeek-V4-Flash-0731 "is the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities." DeepSeek's API change log states that DeepSeek-V4-Flash-0731 "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained" — the gains come from post-training, not a new design. The checkpoint ships with a DSpark speculative decoding module attached (the same model structure as DeepSeek-V4-Flash-DSpark).

DeepSeek reports that DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on the benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available. Of the original preview, DeepSeek said that DeepSeek-V4-Flash's reasoning capabilities closely approach V4-Pro, that it performs on par with V4-Pro on simple agent tasks, and that its smaller parameter size brings faster response times and highly cost-effective pricing. The technical report notes that DeepSeek-V4-Flash-Max (the maximum reasoning effort mode) achieves comparable reasoning performance to the Pro version when given a larger thinking budget, though its smaller parameter scale naturally places it slightly behind on pure knowledge tasks and the most complex agentic workflows.

Architecture

DeepSeek-V4-Flash shares the DeepSeek-V4 series architecture: a Mixture-of-Experts Transformer with a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for long-context efficiency, Manifold-Constrained Hyper-Connections (mHC) strengthening conventional residual connections, and training with the Muon optimizer. Published configuration details for DeepSeek-V4-Flash from the technical report:

• 43 Transformer layers; hidden dimension 4096.
• The first two layers use pure sliding window attention; subsequent layers use CSA and HCA in an interleaved manner (for CSA, attention top-k of 512 KV entries; an additional sliding-window attention branch uses a window size of 128).
• MoE layers are used in all Transformer blocks, with the Hash routing strategy for the first 3 MoE layers. Each MoE layer consists of 1 shared expert and 256 routed experts with an intermediate hidden dimension of 2048 per expert; 6 routed experts are activated for each token.
• Multi-token prediction (MTP) depth is set to 1.
• mHC expansion factor is set to 4.
• The FFNs follow the DeepSeekMoE paradigm with fine-grained routed experts and shared experts and an auxiliary-loss-free load-balancing strategy.

The DeepSeek-V4-Flash checkpoint is published in FP4 + FP8 mixed precision: MoE expert parameters use FP4 precision, while most other parameters use FP8. In the 1M-token context setting, the V4 hybrid attention design requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2 (measured on DeepSeek-V4-Pro).

Training

Per the technical report, both DeepSeek-V4 models were pre-trained on more than 32T diverse and high-quality tokens, followed by a two-stage post-training pipeline: independent cultivation of domain-specific experts (through supervised fine-tuning and reinforcement learning with GRPO), followed by unified model consolidation via on-policy distillation. DeepSeek-V4-Flash-0731 then underwent an additional re-post-training run for the official July 31, 2026 release, targeting agentic and coding capability; architecture and parameter counts are unchanged from the preview.

Reasoning Effort

Per the official model card, the reasoning_effort parameter supports three levels — low, high, and max — which control how much deliberation the model spends before answering. DeepSeek recommends sampling parameters of temperature = 1.0, with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise; for the high and max reasoning effort levels, DeepSeek recommends a maximum output length of 384K tokens. For code-agent tasks, DeepSeek evaluated the model at the max reasoning effort level with temperature = 1.0 and top_p = 0.95.

Benchmarks

The following table is reproduced in full from DeepSeek's official model card for DeepSeek-V4-Flash-0731. All numbers are as published.

Benchmark	DeepSeek-V4-Flash-0731	DeepSeek-V4-Flash (Preview)	DeepSeek-V4-Pro (Preview)	GLM-5.2	Opus-4.8
Terminal Bench 2.1	82.7	61.8	72.1	81.0	85.0
NL2Repo	54.2	39.4	38.5	48.9	69.7
Cybergym	76.7	38.7	52.7	-	83.1
DeepSWE	54.4	7.3	12.8	46.2	58.0
Toolathlon-Verified	70.3	49.7	55.9	59.9	76.2
Agents' Last Exam	25.2	15.8	16.5	23.8	25.7
AutomationBench Public	25.1	10.8	12.8	12.9	27.2
DSBench-FullStack †	68.7	37.0	41.8	61.8	71.6
DSBench-Hard †	59.6	25.8	31.1	54.5	71.7

Notes from the official card:

1. For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.
2. † DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

Base-Model Evaluation (from the DeepSeek-V4 Technical Report)

The DeepSeek-V4-Flash-Base results below are from the shared DeepSeek-V4 series evaluation; the base architecture is unchanged between the preview and the 0731 release.

Benchmark (Metric)	# Shots	DeepSeek-V3.2-Base	DeepSeek-V4-Flash-Base	DeepSeek-V4-Pro-Base
Architecture	-	MoE	MoE	MoE
# Activated Params	-	37B	13B	49B
# Total Params	-	671B	284B	1.6T
World Knowledge
AGIEval (EM)	0-shot	80.1	82.6	83.1
MMLU (EM)	5-shot	87.8	88.7	90.1
MMLU-Redux (EM)	5-shot	87.5	89.4	90.8
MMLU-Pro (EM)	5-shot	65.5	68.3	73.5
MMMLU (EM)	5-shot	87.9	88.8	90.3
C-Eval (EM)	5-shot	90.4	92.1	93.1
CMMLU (EM)	5-shot	88.9	90.4	90.8
MultiLoKo (EM)	5-shot	38.7	42.2	51.1
Simple-QA verified (EM)	25-shot	28.3	30.1	55.2
SuperGPQA (EM)	5-shot	45.0	46.5	53.9
FACTS Parametric (EM)	25-shot	27.1	33.9	62.6
TriviaQA (EM)	5-shot	83.3	82.8	85.6
Language & Reasoning
BBH (EM)	3-shot	87.6	86.9	87.5
DROP (F1)	1-shot	88.2	88.6	88.7
HellaSwag (EM)	0-shot	86.4	85.7	88.0
WinoGrande (EM)	0-shot	78.9	79.5	81.5
CLUEWSC (EM)	5-shot	83.5	82.2	85.2
Code & Math
BigCodeBench (Pass@1)	3-shot	63.9	56.8	59.2
HumanEval (Pass@1)	0-shot	62.8	69.5	76.8
GSM8K (EM)	8-shot	91.1	90.8	92.6
MATH (EM)	4-shot	60.5	57.4	64.5
MGSM (EM)	8-shot	81.3	85.7	84.4
CMath (EM)	3-shot	92.6	93.6	90.9
Long Context
LongBench-V2 (EM)	1-shot	40.2	44.7	51.5

Agentic and Coding Capabilities

The 0731 release is specifically an agentic/coding refresh: DeepSeek reports substantially enhanced agentic capabilities, with the benchmark table above covering terminal work (Terminal Bench 2.1), repository-level code generation (NL2Repo), security tasks (Cybergym), software-engineering agents (DeepSWE, DSBench-FullStack, DSBench-Hard), and tool use (Toolathlon-Verified, AutomationBench Public, Agents' Last Exam). DeepSeek's API change log also states that the official V4-Flash natively supports the Responses API format and is specifically adapted for Codex. The checkpoint's attached DSpark module provides speculative decoding for faster generation.

Safety Notes

DeepSeek's official model card and technical report for the DeepSeek-V4 series do not publish a dedicated safety evaluation, red-teaming, or responsible-use section. The model is released under the MIT License. As with any general-purpose model, outputs should be reviewed before use in high-stakes applications.

On Run BiOS

On the Run BiOS platform, DeepSeek-V4-Flash (the 0731 checkpoint) is available as the serverless model deepseek-v4-flash with a 1,048,576-token context window, reasoning always on with effort levels low, high, and max, and prompt caching for repeated prefixes.