DeepSeek has released and open-sourced its latest lightweight model, V4 Flash, which beats top models such as GLM 5.2 at a fraction of the cost. It is also quite close to Claude Opus 4.8 in several benchmark tests.
It costs 100 times less than Claude Fable 5.
The underlying mixture-of-experts model has 284 billion parameters, with 13 billion activated per token, and supports a one-million-token context window. Hugging Face lists the complete repository at 304 billion parameters because the checkpoint also includes the DSpark speculative decoding module.
DeepSeek released the weights without access restrictions under the MIT licence, allowing commercial and on-premise deployment.
DeepSeek reported that V4-Flash-0731 beat both the V4-Flash preview and V4-Pro preview across every published agent benchmark. It also scored higher than GLM-5.2 in all nine tests.
V4-Flash came close to Claude Opus 4.8 on Terminal Bench 2.1, DeepSWE, Agents’ Last Exam, AutomationBench Public and DSBench-FullStack. Claude maintained larger leads on NL2Repo, Cybergym, Toolathlon-Verified and DSBench-Hard.
However, all results are vendor-reported. DeepSeek ran public code-agent tests with the minimal mode of its unreleased DeepSeek Harness framework, using maximum reasoning, a temperature of 1.0, and a top-p value of 0.95. DSBench-FullStack and DSBench-Hard are internal DeepSeek datasets, so independent results may differ.
DeepSeek charges $0.14 per million uncached input tokens, $0.0028 per million cached input tokens, and $0.28 per million output tokens.
V4-Pro costs $0.435 per million uncached input tokens and $0.87 per million output tokens, making V4-Flash output roughly one-third as expensive.
The V4-Flash API supports up to 2,500 concurrent requests per account, compared with 500 for V4-Pro. It also supports thinking and non-thinking modes, a one-million-token context window, ow and outputs of up to 384,000 tokens.
The API now natively supports the Responses API format and has been adapted for Codex. The July 31 update only applied to V4-Flash. DeepSeek did not update the V4-Pro API or the models used in its web and mobile applications.
DeepSeek has confirmed that the official V4-Pro model is in development but has not announced a release date.
V4-Flash has 43 Transformer layers and uses a mixture-of-experts system in every block. Each MoE layer contains one shared expert and 256 routed experts with an intermediate dimension of 2,048. Six routed experts activate per token, while the first three MoE layers use hash routing. Its multi-token prediction depth is one.
The model combines Compressed Sparse Attention and Heavily Compressed Attention. It also replaces standard residual connections with Manifold-Constrained Hyper-Connections using an expansion factor of four and 20 Sinkhorn-Knopp iterations.
DeepSeek trained the model on more than 32 trillion tokens. It used the Muon optimizer for most parameters and AdamW for embeddings, prediction heads, and RMSNorm weights.
At a one-million-token context, DeepSeek says V4-Flash uses 10 percent of the single-token inference computing operations and 7 percent of the KV cache required by DeepSeek-V3.2. The separate figures of 27 percent of computing operations and 10 percent of KV cache apply to V4-Pro, not V4-Flash.
The checkpoint includes DSpark speculative decoding, which uses a draft component to predict several tokens before the main model verifies them together.
DeepSeek recommends seven speculative tokens with greedy draft sampling. Its DSpark paper reports 60 to 85 percent faster per-user generation than its previous MTP-1 system at the same overall throughput.
Using the hosted API does not require dedicated AI hardware. Self-hosting is more demanding because all model experts must remain in memory, even though only 13 billion parameters activate for each token.
DeepSeek’s vLLM example runs the full model on one node with four Nvidia GB300 GPUs. Unsloth lists the lossless 8-bit GGUF version at 162GB and the 3-bit version at 103GB. The smaller version requires more than 103GB of combined system memory and graphics memory after runtime overhead.
Full self-hosting is therefore more suitable for organisations with multi-GPU servers, inference clusters or high-memory workstations.
The model does not use a standard Jinja chat template. DeepSeek instead provides a encoding folder with tools for converting OpenAI-compatible messages and parsing completed outputs.
V4-Flash offers low, high, and maximum reasoning settings. DeepSeek recommends a temperature of 1.0, with a top-p value of 0.95 for agent tasks and 1.0 for other workloads. High and maximum reasoning modes support outputs of up to 384,000 tokens.
DeepSeek-V4-Flash-0731 delivers major improvements over the preview without changing the underlying architecture. It beats the V4-Pro preview across every agent benchmark published by DeepSeek while charging about one-third of V4-Pro’s output price.
However, developers should test the model on their own workloads because the results remain vendor-reported; some tests used an unreleased evaluation framework, and two benchmarks are internal.
Get the latest tech news, telecom insights, and product launches wherever you prefer.
Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.
Technology and Automotive Specialist covering the latest cars, smartphones, AI breakthroughs, and...
Shares
