DeepSeek Releases V4-Flash-0731, a 304B-Parameter Model Claiming to Beat Its Own Pro Preview on Agentic Benchmarks
DeepSeek has released DeepSeek-V4-Flash-0731, a 304-billion-parameter model that supersedes its earlier preview version with what the company describes as substantially enhanced agentic capabilities. According to DeepSeek's technical report, the model outperforms the larger DeepSeek-V4-Pro (Preview) on several coding and agent benchmarks despite a far smaller activated parameter count.
DeepSeek has published DeepSeek-V4-Flash-0731 on Hugging Face, the official release version of DeepSeek-V4-Flash that replaces the earlier preview build. The model card describes it as carrying "substantially enhanced agentic capabilities" compared to its predecessor.
Key Specs
DeepSeek-V4-Flash-0731 has 304 billion total parameters and is released under the MIT License, with weights stored in Safetensors format across BF16, I64, F32, F8_E4M3, and I8 tensor types. The model shares its architecture with DeepSeek-V4-Flash-DSpark, meaning it ships with an integrated speculative decoding module called DSpark. Pricing is not yet disclosed, and the company has not stated the exact context window size for this specific checkpoint, though the associated paper is titled "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence," suggesting a large-context design goal for the V4 family.
The model supports a reasoning_effort parameter with three settings — low, high, and max — that control how much deliberation the model performs before responding. For high and max effort levels, DeepSeek recommends a maximum output length of 384,000 tokens. Notably, this release does not ship with a standard Jinja chat template; instead, DeepSeek provides a custom Python encoding library to convert OpenAI-compatible message formats into model input strings.
Benchmark Claims
According to DeepSeek's technical report, DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) across nine listed benchmarks despite having a far smaller activated parameter count. Reported scores include:
- Terminal Bench 2.1: 82.7 (vs. 72.1 for V4-Pro Preview, 85.0 for Opus-4.8)
- NL2Repo: 54.2 (vs. 38.5 for V4-Pro Preview)
- Cybergym: 76.7 (vs. 52.7 for V4-Pro Preview)
- DeepSWE: 54.4 (vs. 12.8 for V4-Pro Preview)
- Toolathlon-Verified: 70.3 (vs. 55.9 for V4-Pro Preview)
- Agents' Last Exam: 25.2 (vs. 16.5 for V4-Pro Preview)
- AutomationBench Public: 25.1 (vs. 12.8 for V4-Pro Preview)
- DSBench-FullStack: 68.7 (vs. 41.8 for V4-Pro Preview)
- DSBench-Hard: 59.6 (vs. 31.1 for V4-Pro Preview)
These figures also place the new model ahead of GLM-5.2 on most listed tasks, though behind the comparison model labeled Opus-4.8 on several benchmarks. DeepSeek states that code-agent evaluations used its unreleased "DeepSeek Harness" framework in minimal mode, with max reasoning effort, temperature 1.0, and top_p 0.95 — a configuration the company controls, so independent verification of these results has not yet occurred.
Deployment
DeepSeek provides vLLM instructions for enabling DSpark speculative decoding via a single --speculative-config flag, and documents a reference deployment on a single 4×GB300 node using FP8 KV-cache, expert parallelism, and a custom MoE backend. Local deployment guidance, including weight conversion scripts, is available in the model's inference folder. As of publication, no third-party inference provider has listed hosted access to the model.
What This Means
DeepSeek continues its pattern of shipping technically detailed, self-hosted model releases with permissive MIT licensing rather than restrictive commercial terms, lowering the barrier for developers to run frontier-scale agentic models on their own infrastructure. The reported benchmark gains over DeepSeek's own Pro Preview model — achieved with fewer activated parameters — align with the company's continued emphasis on architectural efficiency over raw scale. However, all benchmark comparisons come from DeepSeek's internal technical report using an unreleased evaluation harness, so the agentic performance claims relative to GLM-5.2 and Opus-4.8 remain unverified until independent testing catches up. Pricing, exact context window figures, and third-party hosting availability are still pending.
Related Articles
NVIDIA Releases Nemotron VoiceChat 11B, an Open Full-Duplex Speech Model with Live Tool Calling
NVIDIA has released NemotronLabs VoiceChat 11B, an 11-billion-parameter end-to-end full-duplex speech model that unifies streaming speech understanding and generation in one architecture. The model claims to be the first open full-duplex system to support live tool calling during natural conversation, with ~450ms turn-taking latency.
Liquid AI Releases LFM2.5-2.6B, a 2.6B-Parameter Agentic Model with 128K Context for On-Device Use
Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model trained on 34 trillion tokens with a 128K context window, built for on-device agentic workloads. The company claims it is competitive with models four times its size on tool use and instruction following.
Liquid AI Releases LFM2.5-2.6B, a 2.6B-Parameter Agent Model for On-Device Deployment
Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model designed to run capable tool-calling agents locally on laptops and phones. The company claims it matches or beats models up to 4x its size on instruction-following and tool-use benchmarks while running under 2.5GB of memory.
Alibaba Unveils Qwen3.8-Max, a 2.4T-Parameter Open-Weight Model for Coding and Agentic Work
Alibaba announced Qwen3.8-Max, a 2.4T-parameter flagship model targeting coding and long-horizon agentic work, with open weights promised for next week alongside Qwen3.8-27B. The model posted strong third-party benchmark results, ranking #4 in Frontend Code Arena and matching Claude Opus 4.7 on the Vals Index at roughly 2.3x lower cost.
Comments
Loading...