DeepSeek V4 Pro or Kimi K2.7 Code — which is the best open-weight coding model?

DeepSeek V4 Pro leads raw coding benchmarks: SWE-bench Verified ~91% (reports range 80.6%-91.2%), HumanEval ~96.4%, MBPP+ ~91.1%, LiveCodeBench 93.5%. Kimi K2.7 Code leads agentic tool use: MCP Mark Verified 81.1 (higher than Claude Opus 4.8) and MCP Atlas 76.0, plus native multimodal input (text/image/video). For pure coding output, DeepSeek V4 Pro wins. For agentic workflows with heavy tool use — especially anything MCP-integrated — Kimi K2.7 Code wins.

What are the specs of DeepSeek V4 Pro and Kimi K2.7 Code?

DeepSeek V4 Pro: 1.6T total parameters, 49B active per token, Mixture-of-Experts, 1M context window, text-only, open weights. DeepSeek V4 Flash: 284B total / 13B active, same 1M context. Kimi K2.7 Code: 1T total parameters, 32B active per token, MoE, 256K context, native multimodal (text + image + video), open weights. Both are true open-weight releases downloadable from Hugging Face (DeepSeek-V4-Pro, Moonshot Kimi-K2.7-Code). Both can be self-hosted with sufficient GPU capacity or served via API.

Which open-weight model is cheaper via API?

DeepSeek V4 Flash is the cheapest of the two by a wide margin — it's positioned as an efficiency tier and typically costs 3-5x less per token than Kimi K2.7 Code via first-party APIs. DeepSeek V4 Pro is also cheaper than Kimi K2.7 Code per token, but the gap narrows on token efficiency: Kimi K2.7 Code cut 'thinking-token' usage by ~30% vs K2.6, closing some of the price gap. For coding-agent workloads that emit lots of tool-call reasoning, measure task-completion cost, not per-token cost.

Do DeepSeek V4 or Kimi K2.7 Code work with Cursor and Claude Code?

Both are accessible via OpenRouter, Together, Fireworks, and other model aggregators that Cursor's Composer and MCP-compatible tools can route to. Cursor supports both via its multi-model configuration. Claude Code is Anthropic-primary but can call external MCP servers; you can build MCP tools backed by DeepSeek V4 or Kimi K2.7 Code for tool-execution roles. Native first-party integration varies: DeepSeek has broader distribution across Chinese cloud platforms; Kimi K2.7 Code is available in Microsoft Azure AI Foundry as of 2026.

Can I self-host DeepSeek V4 Pro or Kimi K2.7 Code?

Yes, both have open weights. DeepSeek V4 Pro (49B active MoE) typically needs 2-4 H100/H200 GPUs for reasonable throughput with vLLM or SGLang. Kimi K2.7 Code (32B active MoE) is somewhat lighter — 1-2 H100/H200 for single-user, 2-4 for multi-user. Both benefit from tensor parallelism and MoE-aware serving frameworks. For local dev on a single 5090 or M3 Ultra Mac: DeepSeek V4 Flash (13B active) is more realistic than either Pro variant. For production self-hosting, budget $8-25K/month in cloud GPU costs vs $500-3000/month via API — self-host only when you have >100M tokens/month or strict data-residency needs.

Quick Answer

DeepSeek V4 Pro vs Kimi K2.7 Code: Best Open-Weight Coding Model? (July 2026)

Published: July 11, 2026

The Setup

Open-weight coding models had a big first half of 2026. Two now lead the class:

DeepSeek V4 Pro — 1.6T total / 49B active MoE, 1M context, text-only, top raw coding scores
Kimi K2.7 Code (Moonshot AI) — 1T total / 32B active MoE, 256K context, native multimodal, top agentic tool-use scores

Both are open-weight (downloadable, self-hostable). Both are frontier-competitive. They optimize for different workloads.

Specs Side-by-Side

Attribute	DeepSeek V4 Pro	Kimi K2.7 Code
Architecture	MoE	MoE
Total parameters	1.6T	1T
Active per token	49B	32B
Context window	1M tokens	256K tokens
Modality	Text only	Text + image + video
License	Open weights	Open weights
Sibling model	V4 Flash (284B / 13B active)	(single tier)

Benchmarks

Benchmark	DeepSeek V4 Pro	Kimi K2.7 Code
SWE-bench Verified	~91.2% (some reports 80.6%)	60.4%
HumanEval	~96.4%	—
MBPP+	~91.1%	—
LiveCodeBench	93.5%	—
Kimi Code Bench v2	—	62.0
Program Bench	—	53.6
MLS Bench Lite	—	35.1
MCP Mark Verified	—	81.1 (>Opus 4.8)
MCP Atlas	—	76.0
Artificial Analysis Intelligence Index	44	42

On raw coding output: DeepSeek V4 Pro is the clear winner. On agentic tool use (especially MCP): Kimi K2.7 Code is the clear winner — even beating Claude Opus 4.8 on MCP Mark Verified.

When to Pick Each

Pick DeepSeek V4 Pro for:

Raw code generation (SWE-bench-style problems)
Long-context codebase understanding (1M window)
Cost-sensitive coding at scale (cheaper than Kimi K2.7 Code per token)
Text-only workflows where multimodal isn’t needed
Self-hosting with strong benchmark ROI

Pick DeepSeek V4 Flash for:

High-throughput cheap coding (efficiency variant at ~13B active)
Local dev on a single high-end GPU
Budget batch tasks
Prototyping

Pick Kimi K2.7 Code for:

MCP-integrated agentic workflows (leads the class)
Multi-step tool use in agent frameworks
Multimodal input (code + screenshots + video walkthroughs)
Long-horizon tasks where its 30% thinking-token reduction (vs K2.6) matters
Workflows already targeting Kimi’s ecosystem or Azure AI Foundry

Cost Considerations

Via API (typical July 2026 pricing on aggregators):

Model	Rough $/MTok input	Rough $/MTok output
DeepSeek V4 Flash	~$0.30	~$1.20
DeepSeek V4 Pro	~$0.60	~$2.40
Kimi K2.7 Code	~$1.20	~$5.00

(Prices vary by provider; check OpenRouter/Together/Fireworks for current rates.)

DeepSeek V4 Flash is decisively the cheapest name-brand open-weight coding model of July 2026.

Self-hosted — GPU budget:

Model	GPUs for single-user	GPUs for prod (10-100 QPS)
DeepSeek V4 Flash (13B active)	1x H100	2-4x H100
Kimi K2.7 Code (32B active)	1-2x H100/H200	2-4x H100/H200
DeepSeek V4 Pro (49B active)	2-4x H100/H200	4-8x H100/H200

Break-even for self-hosting vs API: roughly 100M+ tokens/month, or strict data-residency needs.

The Multimodal Advantage of Kimi K2.7 Code

Kimi K2.7 Code accepts image and video input natively. That opens workflows that DeepSeek V4 (text-only) can’t touch:

Debug from a screenshot of an error dialog
Refactor UI code from a Figma image
Generate code from a video walkthrough of expected behavior
Analyze visual diffs in generated UI

For agentic frameworks where the agent needs to “see” something to act on it, Kimi K2.7 Code is a genuinely different category.

MCP Integration: Why Kimi K2.7 Code Wins There

Kimi K2.7 Code posts MCP Mark Verified 81.1 — higher than Claude Opus 4.8. That’s on tool-invocation accuracy in agentic workflows using the Model Context Protocol.

For teams building on MCP (especially with the 2026-07-28 stateless spec landing later this month), Kimi K2.7 Code is currently the best open-weight model for MCP tool use. DeepSeek V4 Pro doesn’t publish comparable MCP benchmarks.

How They Compare to Closed Frontier Models

Model	SWE-bench Verified	Cost band
DeepSeek V4 Pro	~91.2%	Cheapest name-brand tier
Claude Code (Sonnet 5)	80.9%	$2/$10 (intro)
GPT-5.6 Sol	~85% (est.)	$5/$30
Grok 4.5	~64.7% (SWE-bench Pro, harder variant)	$2/$6
Kimi K2.7 Code	60.4%	Mid

DeepSeek V4 Pro’s SWE-bench Verified score, if it holds up in independent evals, is genuinely at or above closed frontier models — at open weights, at lower per-token cost. That’s the biggest 2026 open-weight story.

Availability

DeepSeek V4 Pro / Flash:

Hugging Face (deepseek-ai/DeepSeek-V4-Pro)
DeepSeek API (deepseek.com)
OpenRouter, Together, Fireworks
Chinese cloud platforms with broad distribution

Kimi K2.7 Code:

Hugging Face (moonshotai/Kimi-K2.7-Code)
Kimi API (platform.kimi.ai)
Microsoft Azure AI Foundry (native)
OpenRouter, Together

The Bottom Line

DeepSeek V4 Pro is the best open-weight coding model for raw code generation in July 2026 — SWE-bench Verified ~91%, 1M context, cheapest tier.

Kimi K2.7 Code is the best open-weight coding model for agentic tool use — MCP Mark Verified 81.1 (beats Opus 4.8), native multimodal input.

They’re not the same product. Pick DeepSeek V4 Pro for coding output, Kimi K2.7 Code for MCP-integrated agents. Many production 2026 stacks use both.