Which open-weight model is best for coding in June 2026?

It depends on the workload. For raw intelligence and long-horizon agentic coding, GLM-5.2 (released June 16, 2026) is the new leader at 51 on the Artificial Analysis Intelligence Index, ahead of DeepSeek V4 Pro max (44) and Kimi K2.6 (43). For MCP tool-heavy agentic loops, Kimi K2.7 Code's MCP Atlas 76.0 and MCP Mark Verified 81.1 scores are still the strongest among open weights. For raw capability per dollar on standard SWE benchmarks, DeepSeek V4 Pro is still the price leader. For a Cursor-style web workspace experience, Kimi Code (launched June 12, 2026 alongside K2.7 Code) is the only first-party agentic surface.

What is GLM-5.2's architectural advantage over DeepSeek V4 Pro?

GLM-5.2 is leaner. At 753B total parameters with 40B active (Mixture-of-Experts), it is smaller than DeepSeek V4 Pro's 1.6T total parameters with around 200B active. The smaller active-parameter footprint means cheaper per-token inference on equivalent hardware, even though DeepSeek V4 Pro can devote more computation to the hardest tasks. GLM-5.2 also has a 1 million token context window versus DeepSeek V4 Pro's standard 128K-200K depending on provider. For long-horizon agentic engineering tasks where context size matters more than raw parameter count, GLM-5.2's architecture is a better fit. For tasks needing the deepest reasoning, DeepSeek V4 Pro's bigger active count still wins.

How does Kimi K2.7 Code's MCP performance compare?

Kimi K2.7 Code, released June 12, 2026 by Moonshot AI, posts MCP Atlas 76.0 and MCP Mark Verified 81.1 — the strongest open-weight scores on the Model Context Protocol tool-use benchmarks as of June 19, 2026. For long agentic loops with many MCP tool calls (Cursor, Claude Code, Cline, Roo Code, OpenCode workflows), this matters. Neither GLM-5.2 nor DeepSeek V4 Pro reports MCP scores at this tier. The honest read: if your agentic stack lives on MCP tool-calling, Kimi K2.7 Code is the open-weight default. If your stack is more about raw model intelligence and long-context coding, GLM-5.2 wins. If your stack is cost-first with standard SWE workloads, DeepSeek V4 Pro is the workhorse.

Which is cheapest to run in production?

DeepSeek V4 Pro is the API price leader at typical sub-$1.00 per million input tokens across providers. GLM-5.2 on OpenRouter is $1.40/$4.40 per million tokens — slightly more expensive than DeepSeek but cheaper than Kimi K2.7 Code at $0.95/$4.00 per million tokens (Moonshot direct API). For self-hosting, the trade-off flips: GLM-5.2's 40B active parameters fits on roughly half the GPU footprint of DeepSeek V4 Pro's ~200B active, making GLM-5.2 substantially cheaper to self-host at scale. Kimi K2.7 Code at ~30B active is the cheapest self-hosted option if you can fit the ~595GB total weights.

Quick Answer

GLM-5.2 vs DeepSeek V4 Pro vs Kimi K2.7 Code: Open-Weight June 2026

Published: June 19, 2026

GLM-5.2 vs DeepSeek V4 Pro vs Kimi K2.7 Code: Open-Weight June 2026

Three frontier open-weight models from three Chinese labs, released in three different June 2026 weeks. GLM-5.2 (Z.ai, June 16), DeepSeek V4 Pro (continuing 2026 leader), and Kimi K2.7 Code (Moonshot, June 12) are now the open-weight options that matter. Here is how they compare for production deployment.

Last verified: June 19, 2026.

TL;DR

GLM-5.2 — new #1 open-weight on Intelligence Index at 51. Best for long-horizon coding, 1M context.
DeepSeek V4 Pro — price-per-capability leader. Largest parameter count. Best for raw SWE workloads.
Kimi K2.7 Code — strongest MCP tool-use scores. Best for agentic MCP-heavy stacks.
Honest pick: Most teams will route across all three based on workload type.

Direct comparison

Spec	GLM-5.2	DeepSeek V4 Pro	Kimi K2.7 Code
Release	June 16, 2026 (open)	Active 2026 leader	June 12, 2026
Lab	Z.ai (Beijing)	DeepSeek (Hangzhou)	Moonshot AI (Beijing)
License	MIT	MIT	Modified MIT
Total parameters	753B	1.6T	1T
Active parameters	40B	~200B	~30B
Context window	1M	128K-200K	256K
Vision input	No	No	Yes (MoonViT)
AA Intelligence Index	51 (#1 open)	44 (max)	~43 (K2.6 baseline)
MCP Atlas	Not reported	Not reported	76.0
MCP Mark Verified	Not reported	Not reported	81.1
OpenRouter input/M	$1.40	<$1.00	varies
OpenRouter output/M	$4.40	<$2.00	varies
Direct API input/M	varies	$0.27-$0.55	$0.95
Direct API output/M	varies	$1.10-$2.19	$4.00
Self-host VRAM	~half DeepSeek	Multi-H100/H200	~595GB weights
First-party workspace	No	No	Kimi Code (kimi.com/code)

When GLM-5.2 wins

Long-horizon agentic coding. 1M context window plus optimization for multi-step engineering loops.
Top open-weight intelligence. 51 on Intelligence Index v4.1 is the highest open score.
Self-host cost efficiency. 40B active parameters fits on fewer GPUs than DeepSeek’s ~200B active.
Front-end coding. Ranked #2 on Code Arena WebDev despite no vision input — second only to Claude Fable 5.

When DeepSeek V4 Pro wins

Pure cost-per-capability. Still the lowest API pricing per task across most providers.
Standard SWE workloads. 1.6T parameter MoE has more raw computational headroom for the hardest single-turn coding problems.
Ecosystem maturity. Available on dozens of inference platforms globally with OpenAI-compatible APIs.
Community fine-tunes. MIT license plus existing community has produced the most derivative variants.

When Kimi K2.7 Code wins

MCP tool-heavy workflows. MCP Atlas 76.0 and MCP Mark Verified 81.1 are the open-weight ceiling.
You need vision input. MoonViT encoder for screenshots, diagrams, scientific figures.
First-party agentic workspace. Kimi Code (kimi.com/code) is the only Cursor-style web workspace among the three.
You self-host with limited GPUs. ~30B active means cheapest per-token inference at scale.

The architecture story

The three labs made three different bets on what frontier open-weight should look like:

Z.ai (GLM-5.2): Lean MoE, biggest context window. Bet on long-horizon agentic depth.
DeepSeek (V4 Pro): Maximum parameter count, lowest price. Bet on capability-per-dollar at the model layer.
Moonshot (K2.7 Code): Mid-size MoE, best MCP tool-use, first-party workspace. Bet on agentic ecosystem.

In June 2026, none of these bets is obviously right. All three are top-10 open-weight models. All three are MIT or near-MIT licensed. All three have multiple commercial inference providers.

Routing recommendations

Workload	First choice	Fallback
Long-horizon agentic coding	GLM-5.2	Kimi K2.7 Code
Cost-sensitive bulk SWE	DeepSeek V4 Pro	GLM-5.2
MCP tool-heavy agents	Kimi K2.7 Code	GLM-5.2
Vision-required coding	Kimi K2.7 Code	(closed: Fable 5)
Web Composer-style UI	Kimi Code workspace	Cursor + GLM-5.2
Self-host with limited VRAM	Kimi K2.7 Code (30B active)	GLM-5.2 (40B active)
1M context required	GLM-5.2	Kimi K2.7 Code (256K)

Why this matters now

With Claude Fable 5’s free Pro/Max access ending June 22, 2026, and credit-based pricing kicking in June 23, every production team using Fable 5 inside Claude Code or Cursor is now doing the routing math. The three models above are the open-weight options that don’t lock you into a single closed-frontier vendor’s pricing schedule.

The pattern that is winning in late June 2026:

Keep Claude Fable 5 (or Opus 4.8) as the quality ceiling for the hardest 10-20% of tasks.
Route 60-80% of agentic coding bulk to one of these three open-weight models based on workload type.
Use OpenRouter or a self-hosted vLLM/SGLang stack to abstract the routing.

The open-weight tier is no longer “good enough for cheap work.” It is “competitive on capability with closed frontier for everything except the absolute hardest tasks.” June 2026 is when that became true.