DeepSeek Releases V4-Flash-0731, a 284B-Parameter Model That Beats Its Own Larger Pro Variant on Agentic Benchmarks

DeepSeek has released DeepSeek-V4-Flash-0731, the official production version of DeepSeek-V4-Flash, superseding the earlier preview build. Unsloth has published GGUF quantizations of the 284-billion-parameter model, making it runnable on local and self-hosted hardware.

What's new

DeepSeek-V4-Flash-0731 shares its architecture with DeepSeek-V4-Flash-DSpark, including an attached speculative decoding module for faster inference. According to DeepSeek, the model shows "substantially enhanced agentic capabilities" compared to the preview release.

The most notable claim from DeepSeek: V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) across every listed benchmark despite activating far fewer parameters. DeepSeek describes the model as "broadly competitive with the strongest proprietary models available."

Benchmark results (as reported by DeepSeek)

Benchmark	V4-Flash-0731	V4-Flash (Preview)	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

These figures come from DeepSeek's own technical report and have not been independently verified. DSBench-FullStack and DSBench-Hard are internal DeepSeek test sets. Code-agent evaluations used the minimal mode of the not-yet-released DeepSeek Harness, at max reasoning effort, temperature 1.0, top_p 0.95.

Opus-4.8 leads on most benchmarks, though DeepSeek-V4-Flash-0731 closes the gap substantially compared to the preview version and edges out GLM-5.2 on several coding and automation tasks.

Availability and quantization

Unsloth has released the model in GGUF format using its Dynamic 2.0 quantization method, which the company claims delivers superior accuracy relative to other quantization approaches. Two sizes are currently available:

• UD-Q4_K_XL: 155GB
• UD-Q8_K_XL: 162GB (described by Unsloth as "full precision, lossless")

Unsloth notes smaller quantizations are still in development. The model can also be run through Unsloth Studio, which exposes toggles for "High" and "Max" thinking modes. The model is licensed under MIT, and no inference providers currently host it.

Pricing for API access has not been disclosed, as DeepSeek has not announced a hosted endpoint for this specific release.

What this means

DeepSeek continues its pattern of releasing large models as open weights under permissive licenses, immediately enabling third parties like Unsloth to produce runnable quantized versions rather than waiting for official hosted access. The claim that a "Flash" variant beats its own "Pro" sibling on agentic benchmarks — if accurate — suggests DeepSeek is prioritizing efficient activation patterns over raw parameter scaling for agentic and coding workloads, a trend also visible in Chinese lab releases from GLM and others. The 155–162GB storage footprint still puts this squarely in enterprise or serious hobbyist hardware territory, not consumer GPUs, so broader adoption will likely wait on the smaller quantizations Unsloth says are coming.