DeepSeek V4 is live on Kyma

DeepSeek shipped V4 as a preview release on April 24, 2026. We rolled it onto Kyma the next day. Two variants are live: deepseek-v4-pro for top-tier reasoning, and deepseek-v4-flash for the cheapest 1M-context model on the platform.

What V4 actually is

In DeepSeek's own words from the V4-Pro model card:

A new series of Mixture-of-Experts language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens.

The architectural story under the hood:

• •Hybrid attention combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). At 1M-token context, V4-Pro uses 27% of V3.2's inference FLOPs and 10% of its KV cache. This is what makes 1M-context-by-default economically real, not just technically possible.
• •Manifold-Constrained Hyper-Connections (mHC) — a residual-connection upgrade for cleaner signal flow at depth.
• •Muon optimizer — DeepSeek's first frontier-scale model trained with it; faster convergence, more stable runs.
• •32T+ pre-training tokens, then a two-stage post-training pipeline that cultivates domain-specific experts and consolidates them via on-policy distillation.

Translation for builders: V4 is not a polish pass on V3. The attention rewrite is what's underneath the 1M context, the V4-Flash price drop, and reasoning-by-default without the cost blowing up.

What changed

V4 is a clean step over V3 across the things builders actually feel:

• •1M context by default on both variants (V3 was 128K). DeepSeek's new sparse-attention design means 1M is the standard service tier, not a degraded long-context mode.
• •Native reasoning baked in. V4-Pro and V4-Flash both spend completion tokens on reasoning_content before emitting content, the same pattern as R1.
• •MoE architecture:
  • •V4-Pro: 1.6T total, 49B active per token.
  • •V4-Flash: 284B total, 13B active per token.
• •MIT license on both repositories and weights.
• •65,536 max output tokens, tool calling, structured outputs, and prompt caching all supported.

V4-Flash is the surprise. DeepSeek positions it as "on par with V4-Pro on simple agent tasks", and the price difference is roughly 12x.

Benchmarks (DeepSeek's own numbers)

These are taken from the DeepSeek-V4-Pro model card on HuggingFace. Vendor-published, not Kyma's measurement. We will publish independent numbers once we have real production traffic.

Bold = leader for that row. The honest reading: V4-Pro is a competitive coding flagship. It tops LiveCodeBench, Codeforces, and Apex Shortlist outright, and it stays within a point or two of the best closed-source models on contest math (HMMT) and GPQA Diamond. SWE-Bench Verified is essentially a four-way tie at the top (80.2-80.8).

It loses on a few rows, and we are showing them on purpose. K2.6 is still ahead on SWE-Bench Pro (the harder, multi-step engineering benchmark) and on tool-augmented HLE. Gemini 3.1 Pro is ahead on raw GPQA Diamond and BrowseComp. If your workload is dominated by very long agent chains or web research, K2.6 or Gemini may still be the right call. deepseek-v4-pro is not a strict superset.

Pricing on Kyma

V4-Flash is the headline on price. At $0.19 input it is cheaper than deepseek-v3 ($0.81 input) while adding reasoning, 1M context, and the V4 family architecture. If you are running bulk agent traffic on V3 today, V4-Flash is a likely net win on both quality and cost.

V4-Pro is the premium tier. For comparison: deepseek-v3 runs $0.81 / $2.30, kimi-k2.6 runs $1.28 / $5.40, and the closed-source flagships you see in the table above are 5-30x more expensive per 1M output token. Your price on Kyma is fixed regardless of which inference backend serves the request, and we fall over to a secondary if the primary degrades (4-layer routing).

How to use it

Easiest path - use an alias instead of a model ID:

from openai import OpenAI

client = OpenAI(base_url="https://kymaapi.com/v1", api_key="ky-...")

response = client.chat.completions.create(
    model="reasoning",  # routes to deepseek-v4-pro for top-tier reasoning
    messages=[{"role": "user", "content": "Walk through the proof, then suggest a tighter bound."}]
)

Pin the exact version when you want reproducibility:

response = client.chat.completions.create(
    model="deepseek-v4-pro",      # or "deepseek-v4-flash"
    messages=[...],
    max_tokens=4096,
)

Heads up on max_tokens. V4-Pro and V4-Flash are reasoning models. They spend completion tokens on reasoning_content before emitting the user-facing content. If you set max_tokens: 50 and the model spends 40 on reasoning, you will get an empty content. Give it at least a few hundred tokens of headroom for any non-trivial query, and several thousand for multi-step agent work.

Upgrading from V3

If you hardcoded "model": "deepseek-v3", you have three choices:

• •Drop-in cheaper: switch to deepseek-v4-flash. Cheaper input ($0.19 vs $0.81), 1M context vs 128K, adds reasoning. The behavior shifts because V4-Flash thinks before answering, so size your max_tokens accordingly.
• •Drop-in stronger: switch to deepseek-v4-pro. Higher quality on coding and reasoning, more expensive than V3, same family ergonomics.
• •Stay on V3: keep it. deepseek-v3 remains a first-class active model on Kyma with the same routing and SLA. We do not deprecate without notice.

The family page at /models/deepseek documents all four (V4-Pro, V4-Flash, R1, V3) so you can compare side by side.

What's next

We are watching V4's production behavior over the next week: route stability, tail latency on the long reasoning traces, and how much of the 1M context window real workloads actually fill. If anything changes in the routing (primary swap, prompt caching enablement, etc.) the model pages stay current.

Want to compare on your own workload? Sign in at kymaapi.com - the free credits on signup cover several million tokens of V4-Flash, which is enough to run a real evaluation before you commit.

Links

• •deepseek-v4-pro on Kyma docs
• •deepseek-v4-flash on Kyma docs
• •DeepSeek V4 preview announcement
• •Model weights on HuggingFace (MIT)
• •Pricing page