DeepSeek-V4-Pro-0813: What's New and How It Stacks Up
DeepSeek-V4-Pro-0813 adds DSpark speculative decoding and beats its preview on coding, agentic, and tool-use benchmarks versus GLM 5.2, Kimi K3, and Opus 4.8.
What is DeepSeek-V4-Pro-0813?
DeepSeek-V4-Pro-0813 is the official, non-preview release of DeepSeek-V4-Pro, an open-weight large language model from DeepSeek AI released under the MIT license. It replaces the earlier DeepSeek-V4-Pro (Preview) build, keeps the same underlying model architecture, and adds a new speculative decoding module called DSpark. The headline change isn’t a bigger model, it’s a more efficient and considerably stronger one, with the biggest gains showing up on coding, agentic, and tool-use benchmarks rather than raw knowledge tests.
TL;DR
- DSpark speculative decoding ships as a built-in module in this release, drafting and verifying tokens from the same checkpoint to speed up inference without a separate draft model.
- Coding benchmarks jumped sharply over the preview: DeepSWE went from 12.8 to 62.7, NL2Repo from 38.5 to 61.5, and Cybergym from 52.7 to 83.3.
- Agentic and tool-use scores improved across the board, including Toolathlon-Verified (55.9 to 74.1) and AutomationBench Public (12.8 to 31.8).
- It’s competitive with proprietary frontier models like Opus 4.8 and Kimi K3 on several benchmarks, though it doesn’t lead every category.
- The model card recommends specific inference settings, including
temperature = 1.0,top_p = 0.95for agentic use, and up to 384K output tokens athighormaxreasoning effort. - It runs on vLLM and SGLang with single-flag DSpark activation, tested on hardware configurations like a 4x GB300 node.
- There’s no Jinja chat template this time; DeepSeek instead ships Python encoding scripts to convert OpenAI-style messages into model input strings.
What changed from the preview version?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The preview version of DeepSeek-V4-Pro established the base architecture, but according to DeepSeek’s own benchmark table, it lagged well behind both proprietary competitors and DeepSeek’s own smaller Flash variant on several agentic tasks. The 0813 release closes that gap dramatically.
Look at DeepSWE, a benchmark tied to software engineering agent tasks: the preview scored 12.8, barely functional. The 0813 release scores 62.7, a jump that suggests the earlier model struggled with the agentic loop (planning, executing, self-correcting) that these tasks demand, while the new version handles it competently. Similar patterns show up in NL2Repo (natural language to repository-level code generation), which climbed from 38.5 to 61.5, and Cybergym, a security/exploit-related benchmark, up from 52.7 to 83.3.
Terminal Bench 2.1, which tests a model’s ability to operate in a terminal environment, went from 72.1 to 87.9. That’s a bigger practical signal than it might look: terminal-use benchmarks tend to punish models that mishandle state, retry logic, or long tool-call chains, so a 15+ point gain points to real improvements in how the model manages multi-step execution, not just better single-shot answers.
How does DSpark speculative decoding work?
Speculative decoding is a standard trick for speeding up LLM inference: a smaller “draft” model proposes several tokens ahead, and the main model verifies them in a single pass, accepting the ones that match what it would have generated anyway. The catch is that it usually requires training and maintaining a separate, smaller draft model alongside the main one.
DSpark, as described in the model card, is DeepSeek’s own speculative decoding module, and it’s built into this same checkpoint rather than shipped as a separate draft model. On vLLM, it’s enabled with a --speculative-config flag specifying "method":"dspark" along with a token count and sampling method (the example configuration uses 7 speculative tokens with greedy draft sampling). On SGLang, it’s a single --speculative-algorithm DSPARK flag, and the documentation explicitly notes you should not point to a separate draft model path, since target and draft weights come from the same checkpoint.
Practically, this matters for anyone self-hosting the model: DSpark should mean faster token generation without the operational overhead of managing two models, though actual speedup will depend on hardware and workload, figures the model card doesn’t quantify directly.
How does it compare to GLM 5.2, Kimi K3, and Opus 4.8?
DeepSeek’s own benchmark table puts V4-Pro-0813 in a competitive but not universally dominant position against three other frontier-tier models: GLM-5.2, Kimi K3, and Anthropic’s Opus 4.8.
On HLE (Humanity’s Last Exam) without tools, V4-Pro-0813 scores 42.7, ahead of GLM-5.2’s 40.5 but behind Kimi K3’s 43.5 and Opus 4.8’s 49.8. With tool use enabled, the gap narrows: V4-Pro-0813 hits 60.0, actually surpassing Opus 4.8’s 57.9 and trailing only slightly behind Kimi K3’s 56.0 in the wrong direction, meaning DeepSeek’s tool-augmented score outpaces both competitors here.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
On coding-heavy benchmarks the picture is mixed but generally strong. Terminal Bench 2.1 has V4-Pro-0813 at 87.9, just behind Kimi K3’s 88.3 and ahead of Opus 4.8’s 85.0. DeepSWE tells a different story: Kimi K3 leads at 67.5, V4-Pro-0813 sits at 62.7, and Opus 4.8 trails at 58.0. On Toolathlon-Verified, a tool-use agent benchmark, Opus 4.8 (76.2) and Kimi K3 (76.5) both edge out V4-Pro-0813’s 74.1, but all three comfortably beat GLM-5.2’s 59.9.
The table also includes a row for “Fable-5 (w/ fallback),” which posts the highest scores in several categories, though the source material doesn’t specify what that model is beyond its inclusion as a comparison point.
The overall pattern: DeepSeek-V4-Pro-0813 isn’t the top scorer on every benchmark, but it’s within a few points of Kimi K3 and Opus 4.8 on most agentic and coding tasks, and it clearly outperforms GLM-5.2 on the majority of listed benchmarks. For an open-weight, MIT-licensed model, landing in that neighborhood against closed frontier models is the more notable story than any single benchmark win.
Is DeepSeek-V4-Pro-0813 worth using for agentic and coding work?
For teams building coding agents or tool-calling workflows, the benchmark improvements over the preview are large enough to matter. A jump from 12.8 to 62.7 on DeepSWE isn’t incremental tuning, it suggests the preview version was genuinely unreliable for agentic coding tasks and the 0813 release addresses that directly. Combined with strong Terminal Bench and Cybergym scores, this looks like a model specifically tuned for the kind of multi-step, tool-using workflows that production coding agents rely on.
The MIT license and open weights are also relevant for teams that want to self-host rather than depend on an API. The model card provides concrete deployment paths through vLLM and SGLang, including DSpark activation, quantization details (the release ships with fp8 weights), and recommended sampling parameters for both agentic (top_p = 0.95) and general (top_p = 1.0) use cases.
The tradeoffs worth noting: this is a large model, distributed across 66 safetensor shards, meaning serious hardware is required for local deployment (the documented reference configuration uses a 4x GB300 node). It also drops the standard Jinja chat template in favor of custom Python encoding scripts, which adds a small integration step for teams used to plugging straight into existing chat-template tooling.
Frequently Asked Questions
What is DSpark speculative decoding?
DSpark is DeepSeek’s speculative decoding method built into the DeepSeek-V4-Pro-0813 checkpoint. It drafts candidate tokens and verifies them using the same model weights, rather than requiring a separate smaller draft model, and is enabled via a single configuration flag in both vLLM and SGLang.
How much better is DeepSeek-V4-Pro-0813 than the preview version?
It shows large gains on coding and agentic benchmarks specifically. DeepSWE rose from 12.8 to 62.7, Cybergym from 52.7 to 83.3, and AutomationBench Public from 12.8 to 31.8, according to DeepSeek’s published benchmark table.
Does DeepSeek-V4-Pro-0813 beat Opus 4.8 or Kimi K3?
It’s competitive but doesn’t lead every benchmark. It outperforms Opus 4.8 on HLE with tools (60.0 vs 57.9) and Terminal Bench 2.1, but trails Kimi K3 on DeepSWE and both Kimi K3 and Opus 4.8 on Toolathlon-Verified.
What license is DeepSeek-V4-Pro-0813 released under?
The model and repository are released under the MIT license, according to the official model card.
What inference settings does DeepSeek recommend?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
DeepSeek recommends temperature = 1.0 with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise. For high or max reasoning effort levels, it recommends a maximum output length of 384K tokens.
