TL;DR: Inkling is Thinking Machines Lab’s first model — a 975B-parameter, 41B-active MoE released open-weights under Apache 2.0 on July 15, 2026. It’s a genuinely strong generalist with a 1M context window and native multimodality, but as a plug-in coding backend for Cursor or Cline it’s outclassed: it trails Claude Sonnet 5 and even open-weight rivals GLM 5.2 and DeepSeek V4 Pro on the coding benchmarks that matter, while its API costs more per token than both.

	Inkling	DeepSeek V4 Pro	Claude Sonnet 5
Best for	Multimodal tasks, fine-tuning base, calibrated answers	Cheapest capable agentic backend	Accuracy-critical refactors, hard bugs
Input / Output per 1M	$1.87 / $4.68 (64K ctx; 50% launch discount live)	$0.435 / $0.87 (valley)	$2 / $10 (intro), $3 / $15 std
SWE-bench Verified	77.6%	80.6%	85.2%
Terminal-Bench 2.1	63.8%	67.9%	80.4%
License	Apache 2.0	MIT	Proprietary
The catch	Priced above cheaper open rivals; 975B is not self-hostable for most	Price doubles during Beijing peak	2–5× pricier per token

Honest take: Inkling is the most important open model release of the summer, but it is not your coding backend. If you want a cheap open-weight backend, use DeepSeek V4 Pro; if you want the best open-weight coding scores, use GLM 5.2; if you want the best coding period, pay for Claude Sonnet 5. Reach for Inkling only when you need multimodal input, calibrated uncertainty, or a permissive base to fine-tune.

Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati — shipped its first model on July 15, 2026, and did something the frontier labs won’t: released the full weights under Apache 2.0. The developer internet lit up, and the obvious question for anyone paying for Cursor Pro or running Cline landed within hours: can I point my editor at this instead? This guide answers that with verified numbers rather than launch-day excitement — the specs, the coding benchmarks, the real API cost, the BYOK setup, and the honest verdict on where Inkling fits.

What Inkling actually is

Inkling is a mixture-of-experts model with 975 billion total parameters that activates roughly 41 billion per token. Thinking Machines trained it from scratch in about nine months on 45 trillion tokens spanning text, images, audio, and video — so it reasons natively across all four modalities, not through a bolted-on vision adapter. It ships with a 1M-token context window and full weights on Hugging Face as thinkingmachines/Inkling under Apache 2.0.

Two design choices set it apart from the open-weight pack:

• Controllable thinking effort. You can dial reasoning up or down at request time, trading latency for depth. For a tool-calling loop this matters — leaving a model in max-reasoning mode inside an agentic harness burns output tokens on every turn.
• Calibrated answers. Thinking Machines tuned Inkling to flag uncertainty rather than confidently guess. For code, an honest “I’m not sure this import path exists” beats a fabricated one, so this is a real, if hard-to-benchmark, advantage.

Crucially, Thinking Machines does not pretend this is a leaderboard-topper. Their own launch language calls Inkling “not the strongest overall model available today, open or closed.” The company is positioning it as a balanced generalist and a fine-tuning base — something organizations customize through Tinker, its model-customization platform — rather than a finished frontier product. Take them at their word; the benchmarks back it up.

The coding numbers, verified

Here’s where the “should I use it in Cursor” question gets answered. Inkling was post-trained on a broad mix that explicitly included agentic code and tool use, and the tool schemas were randomized during training to reduce sensitivity to any single harness — so it does function inside coding agents. The question is how well.

Benchmark	Inkling	For reference
SWE-bench Verified	77.6%	DeepSeek V4 Pro 80.6%, Sonnet 5 85.2%, Nemotron 3 Ultra 70.7%
Terminal-Bench 2.1	63.8%	GLM 5.2 82.7%, Sonnet 5 80.4%, DeepSeek V4 Pro 67.9%
AIME 2026 (math)	97.1%	DeepSeek V4 Pro 96.7%
GPQA Diamond	87.9%	Kimi K2.6 91.1%
IFBench (instruction-following)	79.8%	Kimi K2.6 76.0%, ahead of Fable 5 and GPT-5.6 Sol

Read that top row carefully. On SWE-bench Verified — the 500-problem human-filtered set — Inkling’s 77.6% is respectable and comfortably beats fellow US open model Nemotron 3 Ultra. But it lands behind DeepSeek V4 Pro (80.6%) and well behind Claude Sonnet 5 (85.2%). And SWE-bench Verified is a saturated, partly-contaminated benchmark; the number that actually separates coding models in mid-2026 is the harder, contamination-resistant suite and the agentic runs.

That’s where the second row hurts. On Terminal-Bench 2.1, the benchmark that best mirrors what a coding agent does — read files, run commands, iterate — Inkling scores 63.8%, trailing open-weight leader GLM 5.2 by nearly 19 points and Claude Sonnet 5 by almost 17. Multi-step terminal automation is exactly the workload you’d use Cline or Cursor’s agent mode for, and it’s Inkling’s weakest coding row.

The bright spots are real but off-topic for coding: 97.1% on AIME 2026 (edging out DeepSeek V4 Pro) makes it a strong math reasoner, and its class-leading 79.8% on IFBench means it follows complex instructions more reliably than several frontier models. On a blinded head-to-head like Design Arena’s Agentic Web Dev leaderboard it sits below Claude Sonnet 5 and Fable 5 but above Kimi K2.6 and GPT-5.6 Sol — respectable, not commanding.

One standing caveat: these are launch-window, largely vendor-reported figures. Vendors run tuned harnesses, and the gap between standardized and vendor numbers routinely runs 10–30 points. Treat everything here as a directional ranking, not a promise for your repo.

The price problem

Open-weights usually implies cheap. Inkling breaks that assumption. The API rates Thinking Machines posted, per million tokens:

• At 64K context: $1.87 input, $0.374 cached input, $4.68 output
• At 256K context: $3.74 input, $0.748 cached input, $9.36 output

There’s a 50% launch discount running for a limited time, so effectively ~$0.94/$2.34 at 64K while it lasts. It’s hosted through Together AI, Fireworks, Modal, Databricks, and Baseten — several of which expose OpenAI-compatible endpoints, which is what makes BYOK possible at all.

Now put that next to what you’d actually pay for a coding backend. Take a realistic power-user day: 50 agentic sessions at ~20K output tokens each = 1M output tokens.

Model	Output $/1M	Daily output cost	Notes
DeepSeek V4 Pro	$0.87	$0.87	MIT, valley pricing
Inkling (64K, discounted)	$2.34	$2.34	Apache 2.0, discount expires
GLM 5.2	$4.40	$4.40	MIT open weights
Inkling (64K, list)	$4.68	$4.68	after discount ends
Claude Sonnet 5	$10	$10	intro rate until Aug 31

Even at the launch discount, Inkling costs more than twice what DeepSeek V4 Pro does — and DeepSeek scores higher on both SWE-bench Verified and Terminal-Bench. Once the discount expires, Inkling’s list output price approaches GLM 5.2’s while scoring 19 points lower on Terminal-Bench. There is no coding workload in mid-2026 where Inkling is simultaneously the cheapest and the best option. That’s the crux of the honest verdict.

“But it’s Apache 2.0” doesn’t rescue the argument either. DeepSeek V4 and GLM 5.2 are both MIT-licensed — equally permissive for commercial use and redistribution. Inkling’s license is excellent, but it isn’t a differentiator against the exact rivals that beat it on price and coding quality.

And self-hosting to dodge the API? A 975B model needs on the order of 500GB+ of memory even quantized — this is a multi-GPU-server proposition, not a workstation one. For the reality of what hardware open-weight models actually demand, our sister site runaihome.com maintains a best-local-AI-models-by-VRAM guide; the short version is that Inkling is API-only for all but well-funded teams.

How to wire Inkling into Cursor and Cline (if you still want to)

If you want to try it despite the verdict — it’s a fine generalist for mixed multimodal work — here’s the BYOK path. Because no consumer tool ships Inkling in its native model picker at launch, you route through an OpenAI-compatible provider (Together and Fireworks both offer one).

Cline (VS Code, the cleanest BYOK path):

1. Open the Cline settings → API Provider → OpenAI Compatible.
2. Base URL: your provider’s endpoint (e.g. Together’s https://api.together.xyz/v1).
3. API key: from your provider dashboard.
4. Model ID: the provider’s Inkling slug (built on thinkingmachines/Inkling).
5. Set thinking effort low to medium for agentic loops — max reasoning inflates every tool-call turn’s output cost, and Inkling’s Terminal-Bench score doesn’t justify the spend.

Cursor (BYOK via OpenAI-compatible override):

1. Settings → Models → enable a custom OpenAI-compatible model.
2. Override the base URL to your provider’s endpoint and paste the key.
3. Add the Inkling model ID. Note that Cursor Tab autocomplete stays on Cursor’s own model — BYOK only affects Chat and agent turns, never inline completion.

For the general mechanics of pointing these editors at a custom endpoint, the same wiring covered in our DeepSeek V4 Pro backend guide and GLM 5.2 backend guide applies unchanged — only the base URL, key, and model ID differ.

Who Inkling is actually for

It’s not the developer swapping their Cursor backend to save money — DeepSeek V4 Pro does that job cheaper and better. It’s not the team chasing top open-weight coding scores — that’s GLM 5.2. Inkling earns its place in three narrower cases:

• Multimodal coding-adjacent work. If your task genuinely mixes code with images, audio, or video — parsing screenshots into UI code, reasoning over diagrams — Inkling’s native four-modality training is a real edge that pure coding models lack.
• A fine-tuning base. This is Thinking Machines’ actual pitch. An Apache 2.0 975B generalist you can specialize on your own codebase, via Tinker or your own pipeline, is a strategic asset that a closed API can never be.
• Calibrated, low-hallucination output where a model admitting uncertainty matters more than raw benchmark peaks.

For the broader question of when open weights beat a paid API for a coding team, our open source vs closed source AI coding tools breakdown walks through the trade-offs, and the full price landscape lives in the AI code editor cost comparison. For the open-source ecosystem implications of a lab this prominent shipping permissive weights, aifoss.dev is tracking the Inkling release in depth.

FAQ

Is Inkling free to use? The weights are free under Apache 2.0 — you can download, modify, and commercialize them. Running it via a hosted API is not free: expect ~$1.87/$4.68 per million tokens at 64K context (a 50% launch discount was live at publication). Self-hosting is free of per-token cost but requires serious multi-GPU hardware.

Can Inkling replace Claude Sonnet 5 for coding? No, not on quality. Sonnet 5 leads Inkling on SWE-bench Verified (85.2% vs 77.6%) and Terminal-Bench 2.1 (80.4% vs 63.8%). Inkling is the better pick only if you specifically need open weights plus multimodality.

Does Inkling support tool calling for agentic coding? Yes. It was post-trained on agentic code and tool use with randomized tool schemas to reduce harness sensitivity, so it works inside Cline and Cursor’s agent mode — it’s just not as reliable on multi-step terminal tasks as the benchmark leaders.

Which is the better open-weight coding backend — Inkling, DeepSeek V4 Pro, or GLM 5.2? For pure coding value in mid-2026: DeepSeek V4 Pro on price, GLM 5.2 on benchmark scores. Inkling trails both on coding while costing more, so pick it for multimodal or fine-tuning needs, not as a default coding backend.

Can I run Inkling locally on my own machine? Realistically no. At 975B parameters it needs roughly 500GB+ of memory even quantized — a multi-GPU-server workload, not a desktop one. See runaihome.com for the hardware reality of large open-weight models.

Sources

• Inkling: Our open-weights model — Thinking Machines Lab
• Inkling Model Card — Thinking Machines Lab
• Welcome Inkling by Thinking Machines — Hugging Face
• Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE — MarkTechPost
• Thinking Machines amps up its bet against one-size-fits-all AI with Inkling — TechCrunch
• Thinking Machines open sources Inkling, focused on low cost — VentureBeat
• Inkling benchmarks — OfficeChai
• Inkling AI Model: Benchmarks, Specs & Open Weights — Kingy.ai

Last verified: Jul 17 2026. Pricing and benchmark figures are largely launch-window and vendor-reported; verify current API rates with your hosting provider before committing spend.