DeepSeek V4 Flash 0731 Goes Official: What the Agentic and Coding Gains Mean for Developers

DeepSeek V4 Flash 0731 is the official release of what was an April 2026 preview, retrained specifically for agentic coding workloads at roughly one-third the API price of V4-Pro. Released July 31, 2026, it ships with an MIT license, a 284B-parameter MoE architecture with 13B active parameters, and benchmarks that bring the "flash" tier within striking distance of "pro" tier on tool-use tasks.

What Changed from Preview

The April 2026 preview of V4 Flash handled basic coding tasks but struggled with agentic workflows. Terminal Bench 2.1 sat at 61.8, NL2Repo at 39.4, and DeepSWE at 7.3. The July 0731 revision retrained the model for tool-use and autonomous coding loops. Terminal Bench rose to 82.7, NL2Repo to 54.2, and DeepSWE to 54.4. The DeepSWE jump from 7.3 to 54.4 is the most dramatic single improvement: the model went from failing on software engineering tasks to scoring above GLM-5.2's 46.2.

DeepSeek also adapted the Responses API for Codex-style agent pipelines. The model accepts reasoning_effort levels (low, high, max), supports structured JSON output, and handles tool calls natively. The April preview needed external orchestration for what this version does through a single API call.

Architecture Deep Dive

V4 Flash 0731 uses a Mixture of Experts configuration with 284 billion total parameters and 13 billion active per token. Same architectural class as V4-Pro, but DeepSeek applied different training objectives: more emphasis on tool-calling accuracy, longer terminal sessions, and repo-level code understanding than pure reasoning benchmarks.

The attention mechanism combines Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). CSA compresses older tokens to reduce KV-cache pressure, which matters for the 1M token context window. HCA handles dense reasoning passages where compression loses information. The two systems route different parts of a prompt along different attention paths.

Residual connections use Manifold-Constrained Hyper-Connections (mHC), a design that stabilizes training across the 284B parameter surface without the gradient explosions that typically plague MoE models at this scale. DeepSeek also ships DSark speculative decoding: the model drafts seven tokens greedily before the main network verifies them, cutting inference latency without sacrificing output quality.

For self-hosting, the recommended baseline is a 4xGB300 node. vLLM supports DSark out of the box with expert parallelism and FP4 indexer caching. Unsloth offers GGUF quantizations at approximately 162GB for 8-bit and 103GB for 3-bit. SGLang provides an alternative launch path with its own optimization profiles.

Benchmark Breakdown

The following table compares V4 Flash 0731 against its preview, V4-Pro Preview, GLM-5.2, and Opus-4.8 across six agentic and coding benchmarks.

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
DSBench-FullStack	68.7	37.0	41.8	61.8	71.6

V4 Flash 0731 beats its April preview by over 20 points on Terminal Bench and DeepSWE. It trails Opus-4.8 by 2 to 3 points on the hardest benchmarks but runs at roughly one-third the API cost. Against GLM-5.2, the Flash variant wins on Terminal Bench and DeepSWE but falls behind on NL2Repo and DSBench-FullStack.

API vs Self-Hosting

OrcaRouter lists V4 Flash 0731 pricing at approximately $0.147 per million input tokens (cache miss), $0.295 per million output tokens, and $0.020 per million tokens for cache reads. P50 time-to-first-token is 1.25 seconds; P95 is 1.42 seconds. These numbers place it in the GPT-4o-mini pricing tier while delivering performance closer to GPT-4o on coding tasks.

Self-hosting through vLLM on 4xGB300 GPUs eliminates per-token costs after the hardware investment. The GGUF 3-bit variant at 103GB fits on a single A100-80GB with quantization overhead, though performance degrades measurably compared to the full model. The 8-bit Unsloth variant at 162GB requires two GPUs or a single H100-80GB.

For teams running persistent coding agents, the math favors self-hosting within three to six months of continuous use. For intermittent or low-volume workflows, the API remains cheaper than maintaining GPU infrastructure.

What It Means for Developers

V4 Flash 0731 shifts where the line falls between "cheap and limited" and "expensive and capable" for agentic coding. The April preview was good for chat and simple code. The July revision runs terminal sessions, navigates repositories, and writes multi-file patches at near-pro-tier accuracy. A DeepSWE score of 54.4 means the model can handle real software engineering tasks that the April preview scored 7.3 on.

The MIT license removes the legal ambiguity around the April preview build. Commercial projects can integrate the model without negotiating separate agreements. Combined with the low API pricing and viable self-hosting path, V4 Flash 0731 is the first open agentic model that small teams can deploy without enterprise budget considerations.

The gap to Opus-4.8 on the hardest benchmarks is narrowing but still real. If your workloads involve complex architectural decisions, multi-step debugging across unfamiliar codebases, or heavy formal reasoning, Opus-4.8 still holds an edge. For routine coding tasks, CI automation, bug fixing, and agent-driven development workflows, V4 Flash 0731 delivers 80 to 90 percent of that performance at a fraction of the cost.

Key Takeaways

Terminal Bench 2.1 improved from 61.8 to 82.7 between the April preview and the July 0731 release.
DeepSWE jumped from 7.3 to 54.4, making the model viable for real software engineering tasks.
MIT licensing makes the model available for commercial use without restriction.
API pricing is roughly one-third of V4-Pro, with P50 TTFT under 1.3 seconds.
Self-hosting on 4xGB300 or a single H100 is viable; GGUF quantizations start at 103GB.
Opus-4.8 remains ahead on the hardest benchmarks, but V4 Flash 0731 closes the gap significantly.

FAQ

Is V4 Flash 0731 free to self-host?

Yes. The model is released under the MIT license, which permits commercial and non-commercial use without licensing fees. You only pay for the compute hardware or cloud GPU instances required to run it.

Do I need a GPU cluster to self-host this model?

A single H100-80GB can run the 8-bit GGUF variant (162GB) with some memory overhead management. For the full-precision model, 4xGB300 GPUs are the recommended baseline. Quantized variants reduce memory requirements but may degrade output quality.

How does V4 Flash 0731 compare to GPT-4o?

On coding and agentic benchmarks, V4 Flash 0731 approaches GPT-4o performance at roughly one-third the API cost. It trails Opus-4.8 on the hardest reasoning tasks but outperforms GPT-4o on Terminal Bench and DeepSWE. The actual comparison depends on your specific workload.

Can I use V4 Flash 0731 for commercial projects?

Yes. The MIT license places no restrictions on commercial use, attribution requirements, or derivative works. You can embed the model in proprietary products without sharing your code.

What's the difference between V4 Flash and V4-Pro?

Both share the same 284B MoE architecture, but V4-Pro was trained with different objectives emphasizing reasoning depth and multi-step problem solving. V4 Flash 0731 was tuned for tool use, terminal interaction, and coding workflows. The Pro variant remains ahead on pure reasoning benchmarks; the Flash variant is more cost-effective for agent-driven development.

DeepSeek V4 Flash 0731 Goes Official: What the Agentic and Coding Gains Mean for Developers