DeepSeek V4-Pro 0813 Explained: The Hybrid Attention Architecture Behind Its Quiet GA Launch

DeepSeek shipped an update to its flagship model on August 13, 2026, and for a company that has never been shy about announcing releases, it did so unusually quietly. DeepSeek-V4-Pro-0813 went generally available across the company's app, web interface, and API with a short statement emphasizing "significantly enhanced agent capabilities," a line that was reportedly pulled from the company's own site not long after [8]. That quiet rollout turned out to undersell a genuinely substantial piece of engineering: a hybrid attention architecture, a new way of stabilizing very deep residual stacks, and a 1 million token context window that the underlying research paper says costs a fraction of what the previous generation needed [1].

This post walks through what actually shipped, how the architecture works mechanism by mechanism, and what independent benchmarking organizations found once they got their hands on it, because the honest answer is more interesting than either the hype or the quiet walk-back suggests.

If you would rather show the architecture than describe it, here is a video generation prompt built around the hybrid-attention-meets-stability idea, written for a Wan-style video model:

A tightly woven fabric swatch and a loosely knotted net swatch on a wooden table slowly merge into a single hybrid weave pattern, tight stitches near the center and open strided gaps radiating outward, camera slowly pulling back to reveal a small brass pressure gauge beside the fabric holding perfectly steady in its green zone throughout. Clean scientific still-life motion, warm soft studio lighting, shallow depth of field, no readable text, no logos, no people, smooth steady dolly-out camera move.

Step 1: What Actually Shipped on August 13

DeepSeek-V4-Pro-0813 is the general-availability release of a model that had been circulating in preview since April 2026, when DeepSeek first open-sourced DeepSeek-V4-Pro and its smaller sibling DeepSeek-V4-Flash [2]. The underlying research, published as "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence," describes both as Mixture-of-Experts models: V4-Pro at 1.6 trillion total parameters with 49 billion active per token, and V4-Flash at 284 billion total parameters with 13 billion active per token [1]. Both were pretrained on more than 32 trillion tokens before a post-training pass focused specifically on coding, tool use, and multi-step agentic workflows, which is the capability area the August update was built to sharpen [1].

The headline spec is the context window: 1,048,576 tokens, with a maximum output of up to 384,000 tokens in a single response, released under the MIT license so the weights are freely usable and modifiable [3]. A model that can hold roughly the length of several novels in context at once, and still generate a response longer than most technical books, is not just a bigger number on a spec sheet. It changes what kinds of tasks are even attempted with a single call instead of a chunking pipeline: a full codebase review, an entire legal contract redline, or a long multi-turn agent transcript can now fit in one context without the retrieval and summarization tricks a shorter window forces.

How This Compares to DeepSeek V3.2

The paper frames nearly every efficiency claim relative to DeepSeek-V3.2, the prior generation, and that comparison is worth sitting with for a moment because it explains why V4 needed a new architecture rather than a straightforward scale-up. V3.2 supported a context window of roughly 128,000 tokens [6], meaning V4-Pro's 1,048,576 token window represents roughly an 8x expansion. Under standard full attention, an 8x longer context does not cost 8x more, it costs roughly 64x more compute for the attention mechanism alone, since attention cost scales quadratically with sequence length. Simply scaling V3.2's architecture up to a million tokens would have made serving it prohibitively expensive for anyone outside a handful of well-funded labs, which is exactly the problem the hybrid attention pattern in Step 2 below is built to avoid. Artificial Analysis's own measurement of the jump reflects the result of solving that problem rather than just attempting it: V4-Pro-0813 scores 53 on the Intelligence Index versus 42 for V3.2, a 10 point gain that arrives alongside the context expansion rather than despite it [6]. That pairing, meaningfully more context and meaningfully higher intelligence score in the same release, is the actual engineering achievement here, independent of how the GA checkpoint later compared to competing frontier labs.

Getting there affordably is the hard part, and it is the part the rest of this post focuses on.

Step 2: The Hybrid Attention Architecture

Standard transformer attention scales quadratically with sequence length. At 1 million tokens, naive full attention is not just slow, it is close to computationally impossible at reasonable cost. DeepSeek-V4's answer is a hybrid attention design combining what the paper describes as Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), used together rather than as alternatives [1].

The intuition is straightforward even though the production kernel implementation is not. Most useful information at a given position in a long document is either very close by, the last few paragraphs of context, or falls into a smaller set of genuinely important far-away references, an earlier function definition, a contract clause referenced again later, a fact stated once at the top of a long conversation. A hybrid attention pattern gives every token a dense, full-resolution window over its recent neighbors, the CSA path, while handling everything further back through a much more compressed, strided representation, the HCA path, rather than paying full quadratic cost for the entire history.

Here is a simplified version of what that pattern looks like structurally:

python
/code
import numpy as np

def build_hybrid_attention_mask(seq_len: int, local_window: int, compressed_stride: int) -> np.ndarray:
    """Simplified illustration of a hybrid attention pattern combining a dense
    local window (Compressed Sparse Attention style) with a strided long-range
    path (Heavily Compressed Attention style), the two mechanisms DeepSeek-V4
    combines to keep 1M-token context affordable [1]. Real production kernels
    fuse this into custom CUDA/Triton ops; this version just shows the shape
    of the resulting attention pattern."""
    mask = np.zeros((seq_len, seq_len), dtype=bool)
    for i in range(seq_len):
        # Dense local window: every token attends fully to nearby tokens.
        start = max(0, i - local_window)
        mask[i, start:i + 1] = True
        # Compressed long-range path: every token also attends to a strided
        # set of far-away tokens instead of the full history.
        mask[i, 0:i:compressed_stride] = True
    return mask

mask = build_hybrid_attention_mask(seq_len=32, local_window=4, compressed_stride=6)
dense_full = 32 * 32
actual_edges = mask.sum()
print(f"Full dense attention would need {dense_full} edges")
print(f"Hybrid pattern uses {actual_edges} edges ({actual_edges / dense_full:.1%} of dense)")

Running that snippet on a short sequence makes the shape of the saving obvious: a 32-token sequence with a small local window and a strided long-range path uses a small fraction of the edges full dense attention would require, and that gap widens dramatically as sequence length grows toward a million tokens. The paper's own efficiency claim reflects this at production scale: in the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache that DeepSeek-V3.2 needed for the same task [1]. That KV cache reduction matters as much as the compute saving, since KV cache size is usually the actual memory bottleneck that limits how long a context a serving system can support per GPU.

Step 3: Manifold-Constrained Hyper-Connections

A second architectural piece addresses a problem that gets worse as models get deeper: signal amplification through the residual stream. In a standard transformer, each layer adds its output back onto a running residual signal that flows through the entire network. Hyper-connections, a technique for giving the model more flexible control over how much of each layer's output gets added back in, can improve expressiveness, but an unconstrained version risks the residual signal amplifying wildly across dozens of layers, destabilizing training.

DeepSeek-V4 addresses this with what the paper calls Manifold-Constrained Hyper-Connections (mHC), which restricts the hyper-connection weights to a bounded manifold so the signal cannot blow up no matter how many layers are stacked [1]. One detailed third-party technical breakdown of the paper puts a number on the effect: an unconstrained hyper-connection stack can amplify signal by a factor on the order of 3,000x across the depth of the network, while the manifold-constrained version holds that amplification down to roughly 1.6x [11]. Whether or not that exact figure holds precisely in DeepSeek's own internal measurements, the direction of the claim lines up with what the architecture is explicitly designed to do, and it is a useful way to picture why the mechanism matters: without a constraint like this, a very deep, very wide MoE model becomes progressively harder to train stably.

A simplified numeric illustration of the same idea:

python
/code
def residual_stack_amplification(depth: int, per_layer_gain: float, constrain: bool) -> float:
    """Illustrates why unconstrained residual (hyper-connection) stacks blow up
    over many layers, and why a manifold constraint that caps each layer's
    contribution keeps the total gain bounded. This is a simplified numeric
    stand-in for the mHC mechanism DeepSeek-V4 uses to keep deep residual
    stacks stable [1]."""
    total_gain = 1.0
    for _ in range(depth):
        layer_gain = per_layer_gain
        if constrain:
            # A manifold constraint rescales each layer's contribution so the
            # running product cannot diverge, roughly analogous to projecting
            # the hyper-connection weights back onto a bounded manifold after
            # every update.
            layer_gain = 1 + (layer_gain - 1) / total_gain
        total_gain *= layer_gain
    return total_gain

unconstrained = residual_stack_amplification(depth=60, per_layer_gain=1.15, constrain=False)
constrained = residual_stack_amplification(depth=60, per_layer_gain=1.15, constrain=True)
print(f"Unconstrained 60-layer signal amplification: {unconstrained:,.0f}x")
print(f"Manifold-constrained 60-layer signal amplification: {constrained:.2f}x")

The gap between the unconstrained and constrained numbers in that toy example is the whole point: a residual stack with a tiny amount of unchecked amplification per layer compounds into an enormous, unstable multiplier by the time you reach 60 or more layers, while a manifold-constrained version stays bounded regardless of depth.

Step 4: The Muon Optimizer

The third architectural choice worth understanding is training-time rather than inference-time. DeepSeek-V4 was trained using the Muon optimizer rather than the AdamW optimizer that has been the default for large language model training for years [1]. Muon, short for MomentUm Orthogonalized by Newton-Schulz, was introduced by Keller Jordan in late 2024 as one of the first optimizers to meaningfully outperform AdamW at scale [9]. Where AdamW treats every parameter as an independent scalar, Muon recognizes that most transformer weights are matrices with real geometric structure, and it orthogonalizes the momentum update for those matrices using a fast Newton-Schulz iteration instead of an expensive full SVD decomposition [9].

The practical payoff reported across the broader Muon literature is faster convergence and meaningfully lower memory overhead for optimizer state compared to AdamW [9], which matters enormously at the scale DeepSeek-V4 trains at, over 32 trillion tokens across a 1.6 trillion parameter model [1]. The paper credits Muon specifically with faster convergence and greater training stability compared to AdamW for a model of this size [1], and it is notable that DeepSeek, a lab known for training efficiency going back to V3's disclosed training cost, adopted a genuinely newer optimizer rather than sticking with the safer, better-understood default.

Step 5: Calling V4-Pro Directly

Because DeepSeek exposes an OpenAI-compatible API, integrating V4-Pro into an existing pipeline that already uses the openai-python client requires changing little more than the base URL and model name:

python
/code
# DeepSeek's API is OpenAI-compatible, so the official openai-python client
# works directly against it by pointing base_url at DeepSeek's endpoint [10].
from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-chat",  # routes to the current V4-Pro-0813 GA checkpoint
    messages=[
        {"role": "system", "content": "You are a precise coding assistant."},
        {"role": "user", "content": "Write a Python function that checks if a binary tree is balanced."},
    ],
    temperature=0.2,
    max_tokens=1024,
)

print(response.choices[0].message.content)
print(f"Prompt tokens: {response.usage.prompt_tokens}, completion tokens: {response.usage.completion_tokens}")

One detail worth building into any serious usage is DeepSeek's peak/off-peak pricing schedule, which prices the same tokens differently depending on the time of day in UTC. A simple pre-flight check before kicking off a large batch job:

bash
/code
# DeepSeek prices V4-Pro on a peak/off-peak schedule tied to UTC hours, so a
# batch job scheduled at the wrong time of day can cost 2x more for the same
# tokens. A quick shell check before kicking off a large batch run:
current_utc_hour=$(date -u +%H)
if [ "$current_utc_hour" -ge 1 ] && [ "$current_utc_hour" -lt 4 ]; then
  echo "Peak pricing window (01:00-04:00 UTC): consider delaying non-urgent batch jobs"
elif [ "$current_utc_hour" -ge 6 ] && [ "$current_utc_hour" -lt 10 ]; then
  echo "Peak pricing window (06:00-10:00 UTC): consider delaying non-urgent batch jobs"
else
  echo "Off-peak pricing window: good time for large batch jobs"
fi

What Independent Benchmarks Actually Found

Here is where the story gets more interesting than either the muted announcement or the underlying architecture paper alone would suggest. Two of the more credible independent LLM evaluation organizations, Vals AI and Artificial Analysis, tested V4-Pro-0813 after launch, and their results paint a genuinely mixed picture rather than a clean win or loss.

On SWE-bench Verified, a benchmark measuring whether a model can resolve real GitHub issues end to end, Vals AI ranked DeepSeek V4-Pro-0813 second of 82 models tested, at 96.40%, the highest score of any open-weight model on the board and within 0.60 points of the closed-model leader [4]. It achieved that at a fraction of the cost of the models near the top of the board, roughly $0.02 per test run compared to $1.29 for the leading closed model [4]. Reasoning and legal benchmarks improved sharply too: Vals AI recorded a jump from 10.00% to 49.00% on ProofBench and from 23.08% to 40.87% on Legal Research Bench compared to the April preview checkpoint [3].

Artificial Analysis's broader Intelligence Index, which combines nine separate evaluations covering agentic tasks, scientific reasoning, and knowledge, told a less flattering story. V4-Pro-0813 scored 53 on that composite index, an 8 point improvement over the April preview [5], but Artificial Analysis's own reporting noted the score trailed OpenAI's GPT-5.6 Terra by 4 points and Moonshot AI's Kimi K3 by 7 points, and placed the model 12th overall on the separate Vals Index, behind OpenAI's GPT-5.5 among other frontier systems [8]. Among open-weight models specifically, V4-Pro-0813 still placed second on that same Intelligence Index, behind only Kimi K2.6 [6] [7]. The cost story shifted too: that 8 point improvement over the April preview came with roughly a 3.6x price increase for the GA release [7].

Why Benchmark Organizations Disagree

Anyone who reads more than one source on this release will notice the numbers do not always match, and Terminal-Bench 2.1 is the clearest example: DeepSeek's own release materials pointed to a jump from 72.1 to 87.9 on that benchmark, Artificial Analysis's harness measured V4-Pro closer to 79%, and Vals AI's independent reference implementation scored it at just 54.68%, ranking it 36th of 57 models tested [12]. A 33 point spread on the same named benchmark is not a rounding error, and understanding why it happens is more useful than picking whichever number sounds best.

The short answer is that "Terminal-Bench 2.1" is not one fixed test run in the way a school exam is. Each organization implements its own harness: the exact system prompt given to the model, how many retries a failed tool call gets, what counts as a successful task completion, how strictly output formatting is graded, and even which subset of the benchmark's tasks are run, all vary between a vendor's self-reported number, Artificial Analysis's standardized harness, and Vals AI's own reference implementation. A model that has been tuned, even informally through its post-training data mix, toward the kind of prompting and tool-calling conventions one harness happens to use will score differently on a harness that phrases the same task slightly differently. This is exactly why Vals AI explicitly measures models on a neutral, standardized harness rather than accepting vendor-reported numbers, and why serious technical evaluation of any model release should weight independent, methodology-disclosed benchmarks over a single number in a press release, this post included.

The weaknesses were specific rather than vague. Reporting on the release pointed to real difficulty completing tasks inside a sandboxed terminal environment and generating complex financial models in spreadsheet software, both squarely inside the "agentic" capability area the release announcement had emphasized before that statement was pulled [8]. At the same time, the same coverage noted the model impressing specifically in cybersecurity-adjacent tasks, a narrower but genuinely strong result rather than a broad agentic win [8]. Separately, Artificial Analysis measured V4-Pro leading all open-weight models on GDPval-AA, a real-world economically valuable task benchmark, at a score of 1554 versus 1535 for GLM-5.1 and 1514 for MiniMax-M2.7 [6], which is a very different agentic signal than the sandbox and spreadsheet weaknesses above, and a good reminder that "agentic capability" is not one number.

Put together: this is a model that leads or ties the field on structured, well-specified coding tasks like SWE-bench, leads open-weight models on a broad real-world task benchmark, and does so at a cost other frontier labs cannot match, while genuinely lagging closed frontier models on open-ended reasoning composites and struggling with some of the messier, more unstructured agentic tasks its own release announcement highlighted. That is a far more useful picture than either "DeepSeek beats everyone" or "DeepSeek's update flopped," and it is the kind of nuance that gets lost when a single benchmark number gets pulled out of context.

How V4-Pro Compares to Other Frontier Models

A few things stand out from lining these up. V4-Pro's SWE-bench Verified score genuinely competes with, and by Vals AI's measurement edges out, every closed model except the single leader, and it does it at a cost per test that is not in the same order of magnitude as anything else on this list. Its Intelligence Index score trailing GPT-5.6 Terra and Kimi K3 by mid-single digits is a real gap, but a much smaller one than the pricing gap in the other direction. For a deeper look at two of the models on this list, our breakdowns of Kimi K3's agentic coding capabilities and GPT-5.6's Cerebras-powered inference speed cover the other side of this comparison in more depth.

Production Notes: Pricing, Peak Hours, and Picking Pro vs Flash

V4-Pro's pricing runs on a peak/off-peak schedule rather than a flat rate, which is unusual enough that it is worth planning around deliberately rather than discovering by surprise on a bill. Off-peak input and output pricing is meaningfully cheaper than the peak window, which falls during specific UTC hours on weekdays [10]. For any workload that is not latency-sensitive, a large batch evaluation run, an overnight document processing job, a nightly agent sweep, scheduling around the peak window is close to free savings, which is exactly what the bash snippet in Step 5 above is meant to catch before a job kicks off at the worst possible hour.

Choosing between V4-Pro and the smaller V4-Flash checkpoint comes down to the same tradeoff every large-model-versus-small-model decision does, just with unusually well-documented numbers on both sides this time. V4-Flash trails V4-Pro on most evaluations, but the gap on Artificial Analysis's Intelligence Index was just a single point in Flash's disfavor after the two most recent updates [6], which makes Flash a genuinely reasonable default for high-volume, latency-sensitive workloads where V4-Pro's extra cost is hard to justify for the marginal quality gain. Reserve V4-Pro specifically for tasks that resemble what it is actually strong at based on the benchmark picture above: well-specified coding and issue-resolution tasks, and broad real-world economic tasks, rather than open-ended unstructured agent loops where the sandbox and spreadsheet weaknesses reported after launch are more likely to show up.

Inference speed also plays into that decision. Reported figures put V4-Pro's throughput at roughly 76.8 output tokens per second with a time-to-first-token of about 1.74 seconds in one measurement, both notably faster than Claude Opus 5's reported 56 tokens per second and 32.30 second time-to-first-token in the same comparison [12], though as with the benchmark scores above, exact throughput numbers vary by measurement methodology and reporting date, and are worth re-checking against your own serving setup before treating them as fixed.

Speed and cost at this level are not just an infrastructure detail. A pipeline that generates a script, a set of scene visuals, and a voiceover in one pass, the way Text2Shorts in Miraflow AI turns a single topic into a finished vertical short, only feels fast if every model call inside that pipeline responds quickly and cheaply enough to iterate on. The same logic applies to scanning a long uploaded video for its strongest moments with AI Clipping, where the practical cost of running inference across an entire long-form video, not just a single prompt, is exactly the kind of workload where a model's price-to-performance ratio matters as much as its raw benchmark score.

Making the Most of the 1M-Token Window in Practice

A context window this large changes how you should architect a pipeline around it, not just how much text you can paste in. A few practical points worth applying directly:

• Front-load stable reference material, not just the immediate task. Since KV cache cost for a 1M-token context is only 10% of what V3.2 needed for the same length [1], it is now realistic to keep an entire codebase, style guide, or knowledge base resident in context across many calls in a session, rather than re-retrieving fragments of it per request.
• Watch output length separately from input length. The 384,000 token maximum output is generous but not infinite, and a request that asks for an exhaustive line-by-line audit of a very large input can still hit that ceiling. Structuring a long task as several bounded requests, each targeting a specific section, is more reliable than one open-ended maximal-output request.
• Cache-aware prompting saves real money. DeepSeek's pricing separates cache-hit and cache-miss input token rates, and a workflow that repeatedly sends the same large context prefix, a codebase, a document, a system prompt, benefits directly from structuring requests so that stable content stays in the same position across calls, maximizing how often the cache actually hits.
• Do not assume long-context recall is uniform across the window. Long-context models, V4-Pro included, have historically shown some recall degradation for information buried deep in the middle of a very long context compared to the beginning or end. For anything safety- or correctness-critical, placing the most important instructions near the start or end of the context, rather than the middle, remains good practice even with an architecture built specifically to handle long context efficiently.

Common Mistakes and Misunderstandings

A few patterns show up repeatedly in how people talk about this release.

• Treating a single benchmark number as the whole story. V4-Pro-0813 leads on SWE-bench Verified and GDPval-AA while trailing on Artificial Analysis's broader Intelligence Index. Both are true at once, and neither alone describes the model.
• Assuming "agentic capabilities" means uniformly better at every agent-shaped task. The same release that leads open-weight models on GDPval-AA reportedly struggled with sandboxed terminal tasks and complex spreadsheet generation. Agentic benchmarks measure genuinely different skills.
• Ignoring the price increase between the April preview and the August GA release. An 8 point Intelligence Index gain that comes with a roughly 3.6x price increase changes the cost-effectiveness calculus even though the raw capability number went up.
• Assuming V4-Pro and V4-Flash are simply "big" and "small" versions of the same capability. The gap between them has narrowed enough on general intelligence measures that Flash is a legitimate default for many workloads, not just a fallback for when Pro is too expensive.
• Confusing the Muon optimizer with a change to the model's architecture at inference time. Muon changes how the model was trained, not how it runs; it is a training-time efficiency and stability improvement, not an inference-time feature.

Frequently Asked Questions

Is DeepSeek-V4-Pro-0813 open source? Yes. Both V4-Pro and V4-Flash are released under the MIT license, meaning the weights can be downloaded, modified, and self-hosted without the licensing restrictions some other open-weight models carry [3].

What is the actual difference between the April preview and the August 0813 GA release? The underlying architecture, the 1.6 trillion parameter MoE design with hybrid CSA and HCA attention, is the same. The 0813 release is a retrained checkpoint with a post-training pass specifically focused on coding, tool use, and agentic workflows, which is reflected in the benchmark gains on SWE-bench Verified and GDPval-AA, along with the roughly 3.6x price increase over the preview [7].

Should I use V4-Pro or V4-Flash for a production application? It depends on the task. V4-Pro is the stronger choice for well-specified coding and issue-resolution work where its SWE-bench and GDPval-AA scores apply directly. V4-Flash is a reasonable default for high-volume or latency-sensitive workloads, since the Intelligence Index gap between the two narrowed to a single point after the most recent updates [6].

Why did benchmark scores for the same model vary so much between sources in this post? Different evaluation organizations use different harnesses, system prompts, retry policies, and task subsets, even when testing the same named benchmark. The Terminal-Bench 2.1 spread discussed above, from a vendor-reported 87.9 down to Vals AI's 54.68%, is a direct example of why methodology, not just the final number, matters when comparing models.

Does the Muon optimizer change how the model behaves at inference time? No. Muon is a training-time optimizer choice that affected how the model's weights were learned, not a runtime feature. Once training is complete, a model trained with Muon is called and served exactly like any other transformer checkpoint.

Is a 1 million token context window actually usable, or mostly a marketing number? Based on the architecture's efficiency claims, the reduced KV cache and inference FLOPs specifically target making long context practically affordable rather than just technically possible [1], which is a meaningfully different claim than simply supporting a longer window at high cost. As with any long-context model, recall quality can still vary by position within the window, so validate on your own workload before relying on the full length for correctness-critical tasks.

Conclusion

DeepSeek-V4-Pro-0813 is a genuinely significant piece of systems engineering wrapped in an unusually muted launch. The hybrid CSA and HCA attention pattern, the manifold-constrained hyper-connections keeping a deep residual stack stable, and the Muon optimizer accelerating training all point toward the same underlying goal, making a 1 million token context model with agentic ambitions actually affordable to run and to train, and the paper's own efficiency numbers back that up directly. What the independent benchmarks add is the part a research paper alone cannot tell you: this is a model that dominates on structured coding tasks and broad real-world benchmarks at a cost nothing else on the market matches, while still meaningfully trailing the closed frontier on open-ended reasoning and struggling with some of the messier agentic tasks its own launch statement highlighted before that statement quietly disappeared. Read past the headline number in either direction, and that nuance is exactly what makes this release worth understanding in detail rather than summarizing in one sentence.

References and Sources

[1] DeepSeek-AI. "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence."

[2] DeepSeek. Official announcement of the DeepSeek-V4 Preview release, April 2026.

[3] Vals AI: DeepSeek V4 Pro 0813 model page.

[4] Vals AI: SWE-bench Verified leaderboard.

[5] Artificial Analysis: DeepSeek V4 Pro model page, Intelligence, Performance & Price Analysis.

[6] Artificial Analysis. "DeepSeek is back among the leading open weights models with V4 Pro and V4 Flash."

[7] Artificial Analysis, official results post on DeepSeek V4 Pro 0813's Intelligence Index score and pricing change.

[8] South China Morning Post. "DeepSeek's updated V4 Pro AI model struggles on benchmarks, shines in cybersecurity."

[9] Keller Jordan. "Muon: An optimizer for hidden layers in neural networks."

[10] DeepSeek API Docs, news and pricing updates.

[11] BuildFastWithAI. "DeepSeek V4-Pro Review: Benchmarks, Pricing & Architecture."

[12] CodersEra. "DeepSeek V4-Pro (0813) Is Now GA: Benchmarks, Pricing and the Catch."