Inkling vs Kimi K3 vs GLM-5.2 Open-Weight (Jul 2026)
Inkling vs Kimi K3 vs GLM-5.2 Open-Weight (Jul 2026)
Between June 15 and July 17, 2026, three frontier-class open-weight models shipped: Zhipu GLM-5.2 (mid-June), Thinking Machines Inkling (July 15), and Moonshot AI Kimi K3 (July 17). All three are Apache-family permissive licenses, all three claim frontier-adjacent capability, and all three are running on Hugging Face and hosted APIs today.
The open-weight race is now genuinely competitive with closed models — and for the first time, it has a serious US entry (Inkling) alongside Chinese leaders (Kimi K3, GLM-5.2).
Last verified: July 19, 2026
Head-to-Head Spec Table
| Spec | Inkling | Kimi K3 | GLM-5.2 |
|---|---|---|---|
| Vendor | Thinking Machines Lab | Moonshot AI | Zhipu AI |
| Origin | USA | China | China |
| Release date | July 15, 2026 | July 17, 2026 | Mid-June 2026 |
| License | Apache 2.0 | MIT-family (Moonshot) | Apache 2.0-family |
| Architecture | MoE decoder-only | MoE decoder-only | MoE + dense variants |
| Total parameters | 975B (~1T rounded) | 2.8T | ~250B (largest variant) |
| Active parameters/token | 41B | ~50B (per token, MoE routing) | Variant-dependent |
| Context window | 1M tokens | 1M tokens | 1M tokens (top variant) |
| Modality | Multimodal (text + image) | Multimodal (text + image, native visual) | Multimodal |
| Reasoning mode | Controllable effort dial | Long-horizon reasoning built-in | Reasoning variant available |
| Hosted API price | Not yet published | $3 / $15 per MTok | ~$1.20 / $6 per MTok |
| Self-host BF16 weights size | ~1.9 TB | ~5.6 TB | ~500 GB |
| GGUF (Unsloth) size | 270–285 GB | ~700 GB | ~120–200 GB |
| Intelligence Index rank (July 19) | Behind Kimi K3 | 4th (highest open-weight) | Behind Kimi K3 |
| Arena Frontend Code rank | Not ranked (too new) | #1 (above Claude Fable 5) | Strong but below Kimi K3 |
| Fine-tuning platform | Tinker (Thinking Machines) | Kimi Studio / OpenRouter | Zhipu BigModel |
Where Each One Actually Wins
Kimi K3 — the benchmark champion
Wins on:
- Arena Frontend Code #1 (July 2026), passing Claude Fable 5.
- Long-horizon coding. Beat Fable 5 on 6/14 benchmarks in Moonshot’s launch charts; specifically strong on Terminal Bench 2.1, Program Bench, SWE Marathon, Automation Bench, SpreadsheetBench 2, BrowseComp.
- Deep web research (BrowseComp).
- General intelligence for an open-weight model — 4th on the Artificial Analysis Intelligence Index as of July 19, 2026 (behind closed Fable 5, GPT-5.6 Sol, but ahead of all other open-weight models).
- Cost per task. ~$0.94 for typical eval tasks — similar to GPT-5.6 Sol and about half the price of Claude Opus 4.8.
Loses on:
- Self-hosting difficulty. 2.8T parameters means serious compute is needed — most teams will use hosted APIs.
- US procurement. Chinese origin is a real blocker for US federal, defense, and some EU regulated buyers.
- Fable 5 head-to-head on 8/14 Moonshot benchmarks — Kimi still trails Fable 5 in FrontierSWE, DeepSWE, and even Moonshot’s internal Kimi Code Bench 2.0.
Inkling — the US-origin permissive-license play
Wins on:
- US procurement. First frontier-class open-weight model from a US vendor since Meta’s Muse Spark. Not procurement-blocked in China-restricted deployments.
- Apache 2.0 — genuinely permissive, no community-license carveouts.
- Self-hosting ease. 41B active parameters fit reasonable GPU budgets; consumer-hardware GGUFs via Unsloth at 270-285GB.
- Controllable reasoning effort — a runtime dial for cheap-fast vs slow-smart.
- Fine-tuning integration with Tinker platform (Thinking Machines’ hosted service).
Loses on:
- Raw benchmark scores — trails Kimi K3 on general intelligence.
- Coding benchmarks — not on Arena leaderboards yet.
- Ecosystem depth — one week old; adapter and quantization ecosystem still growing.
GLM-5.2 — the cost-efficient generalist
Wins on:
- Lowest hosted API cost in the group (~$1.20 / $6 per MTok).
- Smaller model sizes available (~15B and ~50B variants, plus the 250B flagship) — real flexibility.
- Strong Chinese-language performance — GLM has always led on Chinese tasks.
- Multiple deployment paths (Zhipu-hosted, third-party providers, self-host).
- Mature ecosystem — GLM-5.x line has been running since Q1 2026, adapters and integrations are stable.
Loses on:
- Raw benchmark scores — trails Kimi K3 on international coding leaderboards.
- US procurement — Chinese origin (same as Kimi K3).
- Frontier reasoning — reasoning variants are OK but not category-leading.
Sub-Questions People Are Asking
Which one is best for coding agents?
For raw coding capability, Kimi K3 — it topped Arena Frontend Code and beat Fable 5 on Moonshot’s Terminal Bench and Program Bench evaluations.
For self-hosted coding agents on modest budgets, Inkling — 41B active parameters is much cheaper to run than Kimi K3’s 2.8T-parameter routing overhead.
For teams in Chinese-language contexts, GLM-5.2 with the coding-specialized variant.
Which one is best for RAG?
All three handle 1M context natively. Kimi K3 has the strongest general intelligence at that scale. GLM-5.2 is cheapest for high-token-count RAG. Inkling offers controllable reasoning-effort, which lets you trade cost vs quality per query.
Which one is best for fine-tuning?
Inkling — Thinking Machines built Tinker as the fine-tuning platform for Inkling. If you’re doing serious customization work, this is the most integrated stack.
GLM-5.2 — cheapest fine-tuning through Zhipu BigModel, and community LoRA / QLoRA tooling is mature.
Kimi K3 — fine-tuning support via Cloudflare Workers AI and OpenRouter; less integrated than the other two, but doable.
What about Claude Fable 5, GPT-5.6 Sol, Grok 4.5, Gemini 3.5 Pro?
Those are closed-weight frontier models. This comparison is specifically about downloadable, permissive-license open-weight models. If closed-weight is acceptable, Claude Fable 5 or GPT-5.6 Sol still lead on raw capability — but at 3-5x the token cost of Kimi K3 or GLM-5.2, and without self-host or fine-tune freedom.
What about Muse Spark (Meta) or Qwen 3.5 (Alibaba)?
Both are also strong open-weight models. Muse Spark has the largest ecosystem (largest community of adapters and fine-tunes). Qwen 3.5 is a strong multilingual generalist under Apache 2.0. Both are worth benchmarking against Kimi K3, Inkling, and GLM-5.2 for specific use cases — this comparison focuses on the July-2026 new releases.
Are Chinese open-weight models actually usable in production?
Yes, for most commercial deployments. All ship under Apache-family licenses with no acceptable-use restrictions that meaningfully differ from Meta’s or Google’s. The blockers are procurement policies (US federal, defense, regulated industries) — not license terms. For a typical SaaS or enterprise team, Kimi K3 and GLM-5.2 are commercially usable exactly like Inkling.
How does model size affect self-hosting cost?
Kimi K3 at 2.8T total parameters (~50B active) requires a serious multi-GPU setup — H100 / H200 clusters. Inkling at 975B / 41B active fits smaller inference deployments (a well-configured 8x A100 or 4x H100 box). GLM-5.2 variants from 15B up to 250B give the most flexibility for budget-constrained self-hosting.
The Real Decision Framework
Choose Kimi K3 if:
- You want the highest-performing open-weight model available today.
- You’re okay with Chinese-origin models (most commercial contexts).
- You’ll use it via hosted APIs (Moonshot, OpenRouter, Cloudflare) — self-hosting the full 2.8T model is expensive.
- Your workload is coding, deep web research, or long-horizon agent tasks.
Choose Inkling if:
- You must avoid Chinese-origin models (US federal, defense, some regulated EU).
- You want easy self-hosting on smaller GPU budgets.
- You’ll fine-tune extensively — Tinker gives you the smoothest customization stack.
- You value the controllable reasoning-effort dial.
Choose GLM-5.2 if:
- You need the lowest per-token cost among frontier open-weight options.
- Your workload has Chinese-language content.
- You want to mix model sizes across your stack (fast small variants + one large flagship).
- You already have Zhipu / GLM-x integrations from Q1-Q2 2026.
Choose all three:
- If you’re building an inference-router pattern. Use GLM-5.2 for cheap high-volume queries, Kimi K3 for hard reasoning, Inkling for regulated / customized workflows.
What This Means for the Market
Open-weight is genuinely competitive with closed-weight as of July 2026. Kimi K3 within striking distance of Claude Fable 5 on the Intelligence Index at ~1/3 the token cost is the story of Q3 2026.
The US has a real open-weight vendor for the first time in the frontier era. Inkling gives US-restricted buyers an option — the market is not “all Chinese open-weight or all closed US models” anymore.
Cost pressure on closed-weight vendors is real. Anthropic, OpenAI, Google, and xAI are all watching Kimi K3’s price-performance ratio. Expect another round of closed-weight price cuts in August-September 2026.
Fine-tuning becomes a real product category. Thinking Machines’ Tinker + Inkling is the first integrated US customization stack. Zhipu, Moonshot, and Alibaba have similar offerings. Enterprise fine-tuning is emerging as a meaningful line-item in AI budgets.
Bottom Line
Kimi K3 has the highest ceiling. Inkling has the US-origin advantage. GLM-5.2 has the best cost-efficiency. All three are Apache-family permissive, all three ship on Hugging Face today, all three are worth benchmarking against your workload.
The five weeks between mid-June and July 17, 2026, will be remembered as the point where open-weight frontier models became commercially competitive with the closed-weight majors. If you haven’t stood up a router that can call any of these three (plus Claude Sonnet 5 and GPT-5.6 for closed frontier), your infrastructure is behind the state of the art.
Sources
- Thinking Machines Inkling announcement: thinkingmachines.ai/news/introducing-inkling
- Moonshot Kimi K3 model card: platform.kimi.ai/docs/guide/kimi-k3-quickstart
- Artificial Analysis Kimi K3 review: artificialanalysis.ai/models/kimi-k3
- VentureBeat Kimi K3 coverage: venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems
- The Decoder Kimi K3 vs Fable 5: the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5