DeepSeekโs V4 family has now moved beyond its original preview stage. DeepSeek V4 Flash entered public beta as V4-Flash-0731 on July 31, while DeepSeek V4 Pro reached GA as V4-Pro-0813 on August 13 across APP, Web and API. The two models retain the V4 familyโs one-million-token context design and MoE architecture, with Flash optimized for efficiency and Pro positioned as the higher-capability reasoning and Agent model.
The Pro GA release materially raises the baseline for any V4.1 forecast. DeepSeek says the GA build brings significantly enhanced Agent capabilities, particularly in production environments, and now gives both V4-Pro and V4-Flash native Responses API support plus low, high and max thinking-effort levels. The current API model version is explicitly listed as DeepSeek-V4-Pro-0813. So where does DeepSeek V4.1 fit now? The Informationโs May report said DeepSeek planned a June V4.1 release with stronger enterprise tooling and multimodal support. That window passed, and the current DeepSeek release log still contains no V4.1 model card or release entry. The safest interpretation is therefore a possible post-GA V4 iteration whose reported goals may still arrive under a V4.1 name or through additional V4 updates.
Confirmed milestones use the DeepSeek change log; the earlier V4.1 June plan came from The Information.
What Is DeepSeek-V4.1?
DeepSeek-V4.1 is best understood as an unconfirmed post-GA V4 iteration, not as an officially released DeepSeek model. The V4 family already has a production Pro checkpoint in V4-Pro-0813, so V4.1 would now represent a mid-cycle capability update rather than the step required to complete the original V4 rollout. DeepSeek has not published a V4.1 model card, technical report, API changelog or benchmark table.
| Question | Current answer | Confidence |
|---|---|---|
| Has DeepSeek officially announced โV4.1โ? | No. The current DeepSeek release log has no V4.1 entry or model card. | High |
| Was a V4.1 release reported? | Yes. The Information reported a June plan with enterprise tools and multimodal support. | Medium |
| Did the reported June window happen? | No separate official V4.1 release appeared during that window. | High |
| What is the latest confirmed V4 milestone? | V4-Flash-0731 entered public beta on July 31 and V4-Pro-0813 reached GA on August 13. | High |
| Could V4.1 still launch as a separate model? | Yes, but the name, specifications and timing remain unconfirmed. | Low/Medium |
Expected DeepSeek-V4.1 Specifications
Because DeepSeek has not published V4.1 specifications, the table below separates confirmed current V4 capabilities from directional expectations. โExpectedโ means a reasonable continuation of V4 or a capability reported by media, not an official V4.1 specification.
| Dimension | V4-Flash-0731 | V4-Pro-0813 | DeepSeek-V4.1 (expected) |
|---|---|---|---|
| Model family | V4 | V4 | Likely V4-family iteration |
| Architecture | V4-derived MoE hybrid attention | V4-derived MoE; GA built on Preview structure | Likely V4-derived; not confirmed |
| Total / active params | 284B / 13B | 1.6T / 49B | Unknown; no reliable public number |
| Context window | 1M | 1M | Likely at least the V4 baseline |
| Max output | 384K API maximum | 384K API maximum | Unknown |
| Thinking effort | Low / high / max | Low / high / max | Expected to preserve or extend current controls |
| Responses API | Supported | Supported | Expected if built on current V4 API stack |
| Native image/file input | Not supported in current Responses API | Not supported in current Responses API | Reported multimodal direction; unconfirmed |
| Enterprise / Agent tooling | Functions, web search, apply_patch / Codex support | Functions, web search, apply_patch / Codex support | Reported area of emphasis; MCP expansion remains unconfirmed |
| Open weights | Yes | Yes, V4-Pro-0813 weights available | Possible for a separate V4.1 checkpoint; unconfirmed |
| Official status | Public beta API release | GA | No official release date or model card |
What Could Be New in DeepSeek-V4.1?
Stronger Agentic Coding Without a Bigger Model
The strongest evidence for what comes after V4 is now the V4-Pro-0813 GA update. DeepSeek says Agent capability improved substantially in production environments, with Terminal Bench 2.1 at 87.9, NL2Repo at 61.5 and DeepSWE at 62.7. V4-Flash-0731 had already shown that re-post-training alone could produce large Agent gains without changing model architecture or size.
If V4.1 is a mid-cycle update, the realistic target is no longer simply โbetter than V4 Preview.โ It would need to preserve or exceed the V4-Pro-0813 Agent baseline while improving long-horizon software execution, repository-level reliability, tool orchestration and production consistency.
Better Tool Use and More Complete Agent APIs
DeepSeekโs current V4 stack already gives both Flash and Pro Responses API support and is specifically adapted for Codex. The Responses API compatibility table supports function calls and server-side web search, while file_search, code_interpreter, computer_use and MCP tool types are still ignored. That makes broader enterprise tool coverage a clearer V4.1 opportunity than simply adding the Responses API itself.
Native MCP support is still a plausible upgrade direction, but it remains unconfirmed. A defensible V4.1 thesis is deeper enterprise and Agent integration on top of an API surface that already supports Responses, Codex-oriented workflows, structured output and tool calls.
Multimodal Input Could Be the Biggest Product-Level Upgrade
The Information reported that V4.1 was planned with multimodal support. This remains the clearest possible product-level differentiator because the current DeepSeek Responses API says image and file inputs are not supported. If a future V4.1 checkpoint adds native visual or file understanding, it would expand DeepSeek into document, screenshot and visual-coding workflows that the current text-only API does not cover.
Until DeepSeek publishes a V4.1 model card or API schema, the exact modalities, resolution limits, audio/video behavior and multimodal benchmarks should remain labeled unknown.
Frontier Performance at DeepSeek-Level Cost
Low inference cost remains one of DeepSeekโs most differentiated strengths, but official pricing changed with the full V4 rollout. The current pricing table uses peak/off-peak rates: V4-Flash cache-miss input is $0.22 off-peak / $0.44 peak and output is $0.66 / $1.32 per 1M tokens; V4-Pro input is $0.66 / $1.32 and output is $1.98 / $3.96. Off-peak rates are half of peak rates.
Expected Benchmark Performance of DeepSeek-V4.1
There are still no official V4.1 benchmark scores. The strongest current V4-family baseline is now DeepSeek-V4-Pro-0813 GA, which supersedes the Pro Preview as the reference point for Agent performance. DeepSeek reports 87.9 on Terminal Bench 2.1, 61.5 on NL2Repo, 83.3 on Cybergym, 62.7 on DeepSWE and 74.1 on Toolathlon-Verified, alongside gains on AutomationBench and DSBench.
Figure 6. DeepSeek-V4-Pro-0813 GA official Agent benchmark values, visualized from the DeepSeek change log.
| Benchmark | V4-Pro-0813 | V4-Flash-0731 | V4-Pro Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 87.9 | 82.7 | 72.1 |
| NL2Repo | 61.5 | 54.2 | 38.5 |
| Cybergym | 83.3 | 76.7 | 52.7 |
| DeepSWE | 62.7 | 54.4 | 12.8 |
| Toolathlon-Verified | 74.1 | 70.3 | 55.9 |
| Agents' Last Exam | 25.7 | 25.2 | 16.5 |
| AutomationBench (Public) | 31.8 | 25.1 | 12.8 |
| DSBench-FullStack | 71.1 | 68.7 | โ |
| DSBench-Hard | 67.2 | 59.6 | โ |
Figure 7. V4-Pro-0813 versus V4-Flash-0731 on common Agent benchmarks. Values come from the DeepSeek change log; harness and reasoning settings can affect cross-model comparisons.
The progression suggests three evidence-based expectations for V4.1. First, a meaningful V4.1 release should at least preserve the V4-Pro-0813 GA baseline, not merely beat the old Preview. Second, DeepSeek is still extracting gains through post-training and Agent optimization rather than only scaling parameter count. Third, the most valuable V4.1 upgrade may be broader multimodal and enterprise-tool capability if core Agent performance is already near the top of the V4 design envelope.
DeepSeek-V4.1 vs V4-Flash vs V4-Pro
| Dimension | V4-Flash-0731 | V4-Pro-0813 | DeepSeek-V4.1 expected |
|---|---|---|---|
| Positioning | Fast, efficient Agent/coding model | Current higher-capability V4 reasoning/Agent model | Possible mid-cycle V4 capability upgrade |
| Scale | 284B total / 13B active | 1.6T total / 49B active | Unknown |
| Context | 1M | 1M | Likely at least 1M if V4-derived |
| Agent coding | Very strong after 0731 post-training | Strongest current V4 official Agent baseline | Expected to meet or exceed Pro GA baseline |
| Thinking effort | Low / high / max | Low / high / max | Expected to preserve or extend controls |
| Responses API | Supported | Supported | Expected if built on current V4 stack |
| Multimodal input | No image/file input in current Responses API | No image/file input in current Responses API | Reported possibility; unconfirmed |
| Official API cost | Lower V4 tier; peak/off-peak pricing | Higher V4 tier; peak/off-peak pricing | Unknown |
| Status | Official API release in public beta | GA | Unannounced by DeepSeek |
What Will DeepSeek-V4.1 Be Used For?
If the reported direction holds, V4.1 would be most compelling where the current V4 family is already strong but still lacks broader product capabilities.
1. Repository-scale coding agents. Long-context code understanding, patch generation, testing loops and terminal execution are the clearest fit given the V4-Pro-0813 and Flash-0731 Agent gains.
2. Enterprise research and document workflows. A 1M context baseline, Responses API support and web/tool integration can support large internal knowledge sets, policy corpora and technical documentation. V4.1 would become more useful here if it adds richer retrieval or file tooling.
3. Cost-sensitive autonomous agents. V4-Flash remains particularly attractive for repeated planning and tool-call loops, while V4-Pro offers a higher-capability tier. V4.1 would be compelling if it preserves this cost discipline.
4. Visual document and screenshot analysis. This becomes a major native use case only if the reported multimodal capability is formally delivered.
5. Private or customized deployments. DeepSeek has released V4-Pro-0813 and V4-Flash-0731 weights, so a separately released V4.1 checkpoint could remain attractive for organizations that need self-hosting or domain adaptation. Open-weight availability for V4.1 itself is not confirmed.
What We Still Don't Know
The biggest risk in writing a โcoming soonโ article is turning plausible extrapolation into fake specification. These points should remain explicitly open until DeepSeek publishes first-party material:
- Whether โDeepSeek-V4.1โ will be the official model name.
- Whether V4.1 is a distinct checkpoint or a separate label for improvements beyond V4-Flash-0731 and V4-Pro-0813.
- Total parameter count, activated parameter count and any architectural changes.
- Exact multimodal inputs and whether image, audio, video or files are supported.
- Native MCP, computer-use, file-search or code-interpreter support.
- Official benchmark scores and evaluation settings.
- API pricing, open-weight availability and commercial terms for a separate V4.1 checkpoint. A firm release date.
Final Thoughts
DeepSeek-V4.1 is interesting precisely because the current V4 cycle shows that DeepSeek can make the family materially stronger without introducing a completely new architecture. The Flash-0731 post-training update delivered large Agent gains, and the Pro-0813 GA release pushed that baseline higher again while adding native Responses API support and more flexible thinking effort.
The most credible V4.1 thesis is therefore not โa mysterious new trillion-parameter model.โ It is a more mature V4: stronger production Agent execution, broader enterprise tooling, potentially native multimodal input, and a tighter balance between Flash efficiency and Pro-level reasoning. The central uncertainty is whether DeepSeek will ship those improvements under the V4.1 name at all.
Developers can already test DeepSeek V4 Flash and DeepSeek V4 Pro through CometAPI, while the DeepSeek V4.1 page can be used to track CometAPI availability if a separate model becomes actionable. CometAPI provides a unified API layer for comparing DeepSeek with other frontier models without rebuilding the integration for each provider.