What are the top open-weight AI models in July 2026?

Three lead the open-weight frontier as of July 20, 2026. DeepSeek V4-Pro (April 24 Preview, GA imminent this week) — 1.6T total / 49B active parameters, MIT license, 1M context. Kimi K3 (July 16 release, weights July 27) — 2.8T total parameters, Stable LatentMoE with 896 experts / 16 active, open weights, 1M context. Inkling (Thinking Machines, Mira Murati's startup, July 2026) — Apache 2.0 license, undisclosed parameter count, frontier-adjacent benchmarks. Plus MiniMax M3 Pro (2.7T open weights) as a fourth notable option.

Which open-weight model has the most permissive license?

DeepSeek V4-Pro under MIT license is the most permissive. MIT allows commercial use without restrictions, no copyleft, no share-alike requirements — you can do anything with the weights including modification, redistribution, and commercial deployment. Inkling under Apache 2.0 is a close second — nearly as permissive as MIT with the addition of patent-grant provisions (arguably better for commercial deployment because it explicitly disclaims patent lawsuits from contributors). Kimi K3's specific license terms will be published at the July 27 release; based on Moonshot's prior releases, expect a modified open-source license with some restrictions. For unrestricted enterprise deployment, DeepSeek V4-Pro (MIT) and Inkling (Apache 2.0) are the safest choices.

Which is cheapest to self-host?

Depends on the effective active-parameter count and quantization support. DeepSeek V4-Pro at 49B active parameters is the most efficient of the three — you can run it on 4-8 H100/H200 GPUs for reasonable throughput. Kimi K3's 2.8T total / 16 active experts per token has higher effective compute needs — expect 8-16 H100/H200 for equivalent throughput. Inkling's specifics are less publicly detailed but reportedly optimized for inference efficiency. Rough monthly costs on rented cloud GPUs: V4-Pro at $1500-3000/month; K3 at $2500-5000/month; Inkling estimated $2000-4000/month. For serious workloads, on-prem H200 pods amortize better than cloud rentals.

Which has the best benchmarks?

Kimi K3 leads on aggregate public leaderboards among open-weight models — #4 overall at 80.96 aggregate score, ahead of DeepSeek V4-Pro Preview (~78) and Inkling (frontier-adjacent, exact score depends on which benchmark). On specific benchmarks: K3 is #1 on Arena.ai Frontend Code. V4-Pro leads on some reasoning tests. Inkling has strong performance on academic benchmarks. The DeepSeek V4 GA release (imminent this week) is expected to close some gaps and may shift the aggregate leader position between V4-Pro and K3.

Should I wait for one specific release or use what's available now?

Depends on your timeline. DeepSeek V4 GA is imminent (days away from July 20 as of writing). Kimi K3 weights drop July 27. If you can wait a week: benchmark both against your workload, then pick. If you need something running this week: V4 Preview weights work, Kimi K3 API is available now. For production infrastructure: wait for GA / weights release before committing. For evaluation: start now with what's available and re-evaluate at end of July.

Quick Answer

DeepSeek V4 vs Kimi K3 vs Inkling: Open-Weight (Jul 2026)

Published: July 20, 2026

DeepSeek V4 vs Kimi K3 vs Inkling: Open-Weight (Jul 2026)

Three frontier-adjacent open-weight models, three different bets on how open AI should look. As of July 20, 2026:

DeepSeek V4-Pro — MIT-licensed, 1.6T params, GA release imminent this week
Kimi K3 — 2.8T-param frontier contender, weights dropping July 27
Inkling — Apache 2.0, Thinking Machines (Mira Murati’s startup), frontier-adjacent

Together they represent the strongest open-weight frontier month in AI history. This comparison covers license, benchmarks, self-hosting economics, and ecosystem considerations for enterprise decision-making.

Last verified: July 20, 2026

Head-to-Head Table

Feature	DeepSeek V4-Pro	Kimi K3	Inkling
Developer	DeepSeek	Moonshot AI	Thinking Machines
Founder	Liang Wenfeng	Yang Zhilin	Mira Murati (ex-OpenAI CTO)
Country	China	China	United States
Preview / release date	April 24, 2026 (V4 Preview); GA imminent	July 16, 2026 (API); July 27, 2026 (weights)	July 2026
Parameters (total)	1.6 trillion	2.8 trillion	Undisclosed
Parameters (active)	49 billion	~50-100B effective (16 of 896 experts active)	Undisclosed
License	MIT	Open weights (specific terms at July 27)	Apache 2.0
Context window	1M tokens	1M tokens	Undisclosed but frontier-competitive
Attention mechanism	Compressed Sparse Attention (CSA)	Kimi Delta Attention (KDA)	Undisclosed
API pricing	~$0.30 / $1.20 per MTok (expected, V4 pricing)	$3 / $15 per MTok	Not yet public API
Aggregate leaderboard	~#7-8 (V4 Preview)	#4 (80.96)	Frontier-adjacent
Strongest at	Reasoning benchmarks, long context, extreme cost efficiency	Frontier-adjacent overall, #1 Frontend Code arena	Academic benchmarks, agent workflows
Self-host GPU need (est.)	4-8 H100/H200	8-16 H100/H200	4-12 H100/H200 (est.)
Weights on HuggingFace	Yes (V4 Preview)	Coming July 27	Yes

What Each Model Is

DeepSeek V4-Pro — The MIT-Licensed Efficiency Leader

Positioning: DeepSeek’s flagship open-weight model. V4 Preview shipped April 24, 2026 under MIT license — the most permissive license in mainstream AI. GA release is imminent (within days of July 20) after DeepSeek committed to a mid-July target and multiple industry publications confirm the launch window is closing this week.

Key specs:

1.6T total parameters, 49B active — mixture-of-experts architecture keeps inference cost tractable.
Compressed Sparse Attention (CSA) — DeepSeek’s novel long-context attention mechanism.
1M-token default context — matches Kimi K3, exceeds Sol’s 400K.
MIT license — unlimited commercial use, no copyleft, no restrictions.

Ecosystem signal: DeepSeek raised $7.4B in its first external funding round in July 2026, aimed at global expansion. This is the strongest signal yet that DeepSeek is committing to the open-weight strategy long-term.

Best for: Enterprise self-hosting for compliance / cost, large-scale inference workloads, applications where MIT license simplifies legal review.

Kimi K3 — The Benchmark-Leading Open-Weight Model

Positioning: Moonshot AI’s frontier-adjacent bet. Released via API on July 16, 2026; open weights drop July 27, 2026. #4 on public benchmark leaderboards behind only proprietary Fable 5, Sol, and Gemini 3.5 Pro.

Key specs:

2.8 trillion total parameters — largest of the three.
Stable LatentMoE with 896 experts, 16 active per token — high specialization, tractable inference.
Kimi Delta Attention (KDA) — hybrid linear + Attention Residuals for long-context efficiency.
Native multimodal — image + text handling.
1M-token context.
MXFP4 quantization ready — dramatically reduced memory footprint for deployment.

Ecosystem signal: #1 on Arena.ai’s Frontend Code arena — the best coding benchmark in one specific arena, ahead of both proprietary Sol and Fable 5. Strong demonstration that open-weight can lead specific verticals.

Best for: Coding workflows (especially frontend), general-purpose applications where benchmark leadership matters, self-hosting for cost + compliance in Q4 2026.

Inkling — The Thinking Machines Open-Weight Play

Positioning: Thinking Machines is Mira Murati’s post-OpenAI startup — she’s the former OpenAI CTO. Inkling is the company’s first open-weight release, positioned as Apache 2.0 to signal maximum enterprise-friendliness.

Key specs:

Apache 2.0 license — near-MIT permissiveness with patent-grant provisions (arguably better for enterprise commercial deployment).
Frontier-adjacent benchmarks — competitive with V4-Pro Preview on academic tests.
Parameter count undisclosed publicly (as of July 20, 2026).
US-origin — different geopolitical positioning than Chinese-origin V4 and K3.

Ecosystem signal: Mira Murati’s involvement is the story. Her post-OpenAI positioning as a US-origin, open-weight competitor to closed OpenAI is deliberate. Inkling is Thinking Machines’ opening statement.

Best for: Enterprise deployments where US-origin matters, applications that value Apache 2.0’s patent-grant provisions, users prioritizing Mira Murati’s track record over pure benchmark leadership.

Head-to-Head on Key Dimensions

License Permissiveness

Model	License	Commercial use	Modification	Redistribution	Notes
DeepSeek V4-Pro	MIT	✓	✓	✓	Most permissive; no restrictions
Inkling	Apache 2.0	✓	✓	✓	Near-MIT with patent grants; strong for commercial
Kimi K3	Open weights (specific terms at July 27)	Likely ✓	Likely ✓	Likely with attribution	Modified license; review terms carefully

Winner: DeepSeek V4-Pro (MIT). Runner-up: Inkling (Apache 2.0). Kimi K3 competitive pending specific license terms.

Benchmark Performance

Model	Aggregate score	Best specific benchmark
Kimi K3	80.96 (#4 overall)	#1 Arena.ai Frontend Code
DeepSeek V4-Pro Preview	~78	Strong on reasoning, long-context
Inkling	Frontier-adjacent (varies)	Academic benchmarks

Winner: Kimi K3 on aggregate. DeepSeek V4-Pro GA release may shift this — expected refined weights close gaps.

API + Access Availability

Model	API available today	Weights available today	Enterprise via cloud
DeepSeek V4-Pro	✓ (V4 Preview)	✓ (V4 Preview)	Via DeepSeek + third-party
Kimi K3	✓	Coming July 27	Cloudflare Workers AI, OpenRouter
Inkling	Limited	✓	Third-party hosts emerging

Winner: DeepSeek V4-Pro — most mature availability. Kimi K3 close second — API available, weights imminent. Inkling less broadly available.

Self-Hosting Economics

Rough monthly self-hosted cost estimates (medium-throughput deployment on cloud H100/H200s):

Model	Estimated monthly cost	Effective active params	Notes
DeepSeek V4-Pro	$1500-3000	49B	Best efficiency; MoE keeps inference cheap
Inkling	$2000-4000 (est.)	Undisclosed	Details pending broader deployment
Kimi K3	$2500-5000	Higher effective compute	2.8T total demands more infrastructure

Winner: DeepSeek V4-Pro — lowest self-hosting compute cost. Kimi K3 more expensive but wins on capability-per-dollar for high-capability workloads.

Ecosystem + Community

Model	HuggingFace ecosystem	Fine-tuning support	Community distills
DeepSeek V4-Pro	Strong (multiple V3/V4 forks)	Yes	Extensive (100+ derivatives on V3)
Kimi K3	Growing (July 27 weight release)	Expected yes	Nascent
Inkling	Emerging (Thinking Machines’ first release)	Yes	Nascent

Winner: DeepSeek V4-Pro by ecosystem maturity — DeepSeek V3 had massive community adoption; V4-Pro inherits that.

Real Use Case Comparisons

Use Case 1: Enterprise Self-Hosting for Compliance

Winner: DeepSeek V4-Pro (MIT license). Zero-restriction commercial deployment; strongest self-hosting economics. Runner-up: Inkling (Apache 2.0). Runner-up US-origin option.

Use Case 2: Cost-Optimized Large-Scale Inference

Winner: DeepSeek V4-Pro — API pricing extrapolation ~$0.30/$1.20 per MTok undercuts Kimi K3 by ~10x and Sol by ~15x. Runner-up: Kimi K3 at $3/$15 per MTok.

Use Case 3: Frontend Code Generation

Winner: Kimi K3 — #1 on Arena.ai Frontend Code arena. No competition here among open-weight models.

Use Case 4: Fine-Tuning for Domain-Specific Model

Winner: DeepSeek V4-Pro — most mature fine-tuning tooling and community examples. Runner-up: Kimi K3 post-July 27 as tooling matures.

Use Case 5: US-Origin Preference (Compliance / Geopolitics)

Winner: Inkling — the only US-origin option of the three. Note: open weights + on-prem deployment mitigates most geopolitical concerns with V4-Pro and K3 (data stays in your infrastructure).

Use Case 6: Long-Context Deep Analysis (500K+ tokens)

Tie: V4-Pro or K3 — both 1M context. V4-Pro slightly more efficient at long context due to CSA architecture; K3 has more raw capability. Test both on your workload.

Use Case 7: Academic Research / Reproducibility

Winner: Inkling (Apache 2.0) for maximum reuse rights. Runner-up: DeepSeek V4-Pro (MIT).

Use Case 8: Latest-Capability at Any Cost

Winner: Kimi K3 — highest aggregate benchmark score of the three. Watch: V4 GA release could shift this within days.

The Real Decision Framework

Pick DeepSeek V4-Pro if:

MIT license simplifies legal review.
Self-hosting economics dominate.
You want to inherit DeepSeek V3’s mature ecosystem.
Cheapest API pricing extrapolation.

Pick Kimi K3 if:

Frontend code generation is core.
You want highest aggregate benchmark open-weight model.
2.8T param scale matters for your workload.
Comfortable with Chinese-origin model post-open-weights.

Pick Inkling if:

US-origin matters for compliance.
Apache 2.0’s patent-grant provisions matter.
Mira Murati’s post-OpenAI track record is meaningful to you.
You’re okay with less mature ecosystem than V4-Pro.

Use two or three in parallel:

DeepSeek V4-Pro for high-volume inference + K3 for hardest coding tasks.
Inkling for US-origin regulatory workloads + V4-Pro for cost-optimized batch work.
K3 for benchmark-critical tasks + V4-Pro for cost-critical tasks.

Enterprises with $50K+ AI budgets frequently run two of the three for workload-appropriate routing.

The Bigger Picture: Q3 2026 Open-Weight Consolidation

Three frontier-adjacent open-weight releases in a two-week window is unprecedented. Combined with:

MiniMax M3 Pro (2.7T open weights, released mid-July)
GLM-5.2 (Zhipu, open weights, below frontier)
openPangu 2.0 (Huawei, open weights, below frontier)

The July 2026 open-weight wave establishes that frontier-adjacent AI capability is no longer proprietary-only. Enterprises have real strategic choice on how to build their AI stacks — and open-weight is now a first-class option, not a fallback.

Expect:

Consolidation to 3-5 dominant open-weight models by end of 2026 (V4-Pro, K3, Inkling likely among them).
Community distillations proliferating — 100+ specialized variants on HuggingFace.
Enterprise self-hosting infrastructure buildouts in Q4 2026 as compliance-motivated deployments accelerate.
Proprietary model pricing pressure — Sol and Fable 5 will need to justify their premium over $3/$15 K3 API and effectively-free self-hosted V4-Pro.

Bottom Line

No universal winner in July 2026 — each open-weight model wins specific categories.

DeepSeek V4-Pro — best MIT license, most mature ecosystem, cheapest self-hosting economics.
Kimi K3 — best benchmarks, best frontend code, largest parameter count.
Inkling — best Apache 2.0 license, best US-origin option, backed by Mira Murati’s track record.

For most enterprises evaluating this week: wait for DeepSeek V4 GA (imminent, days away) and Kimi K3 weights (July 27). Benchmark both against your workload, then pick primary + fallback. Don’t lock in a single-model strategy — the open-weight frontier is too fluid right now.

For developers exploring: try Kimi K3 API today ($3/$15 per MTok is cheap enough for real experimentation). Move to self-hosting evaluation after July 27 weights release.

For strategy leads: the open-weight frontier is real capability in 2026. Building AI stack strategy without considering V4-Pro, K3, and Inkling as first-class options is a mistake.

Sources

DeepSeek Wikipedia (V4 details): en.wikipedia.org/wiki/DeepSeek
Kimi K3 official page: moonshot.ai
Digital Applied July 2026 open-weight wave analysis: digitalapplied.com/blog/open-weight-model-wave-july-2026-momentum-tracker
Simon Willison Kimi K3 analysis: simonwillison.net/2026/Jul/16/kimi-k3