Inkling review: Mira Murati's open-weight model, what fine-tuning costs, and when to pick it over Claude API

A 975-billion-parameter open-weight model from Mira Murati's Thinking Machines Lab landed on Hugging Face on July 15, and the API to fine-tune it went live the same day (Thinking Machines launch post). Here is the decision that matters before you spend a week evaluating it. Inkling is not the model you reach for when you want the best coding or reasoning available right now. It trails Claude Fable 5 and GPT-5.6 Sol on nearly every eval Thinking Machines published, and on hard coding it trails Chinese open-weight releases like GLM 5.2 too. What it sells instead is ownership: the full weights, a fine-tuning API built to adapt the model to your own data, and inference at roughly a tenth of frontier API token prices. For most teams without a specific fine-tuning use case, a hosted Claude or GPT plan is cheaper to run and better this quarter. For a team with proprietary data and real inference volume, the math flips, and this review is about where that line sits.

What Inkling is and what it is for

Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active per token, pretrained on 45 trillion tokens of text, images, audio, and video, per the launch post. It reasons natively over text, images, and audio, and carries a context window up to 1M tokens. The architecture largely follows DeepSeek-V3: 256 routed experts plus 2 shared, 6 active per token, a sigmoid router, and sliding-window attention interleaved with global layers at a 5:1 ratio.

Read the design goal twice, because it is the whole pitch. Thinking Machines built Inkling as a broad, balanced base optimized for fine-tuning, not for topping a leaderboard. The post says it flat out: "Inkling is not the strongest overall model available today, open or closed." The bet, per TechCrunch, is that a model an organization can adapt for itself beats the one-size-fits-all models the big labs sell.

Two features carry that thesis. Controllable thinking effort sets how hard the model works per request, from a fast, cheap effort=0.2 to a thorough effort=0.99. And a preview of Inkling-Small, a 276B model with 12B active, matches or exceeds its larger sibling on many benchmarks at lower cost and latency, per the launch post. You can fine-tune on Tinker, hit the base model through APIs on OpenRouter, TogetherAI, Fireworks, Modal, Databricks, and Baseten, or pull the weights from Hugging Face and run them yourself.

What the benchmarks say

Sort the scoreboard into three buckets, all from the benchmark table in the launch post (Inkling run at effort 0.99), fetched 2026-07-19.

Where it holds its own: instruction following and general reasoning. Inkling posts 79.8% on IFBench, ahead of every frontier closed model in the table, including Claude Fable 5 at 63.5% and GPT-5.6 Sol at 72.7% (per the launch post). On AIME 2026 it hits 97.1%, and on GPQA Diamond it lands 87.2%. Competitive numbers for an open-weight base, and the reason the model is worth fine-tuning at all.

Where it trails other open-weight releases: hard coding. Inkling scores 54.3% on SWEBench Pro against GLM 5.2's 62.1%, and 63.8% on Terminal Bench 2.1 while Kimi K2.6 reaches 71.3% and GLM 5.2 reaches 82.7% (per the launch post). The Decoder's read captures the shape: Inkling leads US open-weight labs, topping Nvidia's Nemotron 3 Ultra on the Artificial Analysis Intelligence Index at 41 to 38, but it trails the Chinese releases on coding and text (The Decoder).

Where it lags frontier closed models by a wide margin: the hardest reasoning and factuality. On Humanity's Last Exam, text only, Inkling scores 29.7% against Claude Fable 5's 53.3%; on SWEBench Verified it takes 77.6% while Fable 5 takes 95.0% (per the launch post). Factuality is the softest spot of all, at 43.9% on SimpleQA Verified against 68.3% for Fable 5. Fine-tuning can close a domain gap; it does not turn a mid factuality score into a frontier one.

Calibrate it in one line. If your task is instruction following, structured general reasoning, or a narrow domain you plan to fine-tune, Inkling's ceiling is high enough to matter. If your task is frontier coding or hard multi-step reasoning out of the box, Claude Fable 5 or GPT-5.6 Sol beats a fine-tuned Inkling for most teams, and on the coding evals it is not close.

What fine-tuning on Tinker costs

Tinker bills fine-tuning across three meters, all per million tokens: prefill (your input), sample (generation), and train (the gradient update). A checkpoint stores at $0.10 per GB-month, and cached prefill gets an 80% discount, per the Tinker pricing docs, fetched 2026-07-19.

Price in one caveat first. As of that fetch, the Inkling rows still carry a "Limited-time 50% discount" label, and the launch post flagged the discount as temporary. Treat the numbers below as a floor that roughly doubles when the promotion lifts.

Meter (per 1M tokens, 64K context)	Fine-tuned Inkling on Tinker	Claude Fable 5 API	Source
Input / prefill	$1.87 ($0.37 cached)	$10.00	Tinker / Anthropic
Output / sample	$4.68	$50.00	Tinker / Anthropic
Train (one-time gradient update)	$5.61	not applicable	Tinker
Checkpoint storage	$0.10 / GB-month	not applicable	Tinker

Now the worked example. Say you fine-tune on a 50,000-token dataset for 10 epochs, then serve the tuned model at a support-bot mix of 75% input and 25% output.

Training barely registers. 50K tokens across 10 epochs is 500K train tokens, and at the $5.61-per-million train rate that is about $2.81, one time. A LoRA checkpoint is small; a few GB at $0.10 per GB-month rounds to pocket change (rates per the Tinker docs, fetched 2026-07-19).

Serving is where the decision lives. Run 10 million tokens a month and Inkling on Tinker costs about $14 in prefill (7.5M at $1.87) plus about $12 in sample (2.5M at $4.68), so roughly $26 a month. The same 10M tokens on Claude Fable 5, at $10 input and $50 output per million (per Anthropic's docs), runs $75 plus $125, or $200 a month.

Here is the crossover, and it is not what the "does fine-tuning pay off" question usually assumes. The training cost is trivial and amortizes in a day. The gap that matters is the per-token serving rate, where Inkling on Tinker comes in about eight times cheaper on an input-heavy workload. Even after the launch discount lifts and the rates double, it stays roughly four times cheaper than Fable 5. So the economic question is never whether the fine-tune pays for itself. It is whether Inkling's quality clears your task and whether you can carry the engineering cost of owning a fine-tuned model in production. On price alone, Inkling wins from month one at any real volume.

When to pick Inkling over a hosted API

Pick Inkling when:

• You have proprietary domain data that measurably shifts model behavior, and a fine-tune on it beats a general model on your task. A Bridgewater Associates project cited by TechCrunch hit 84.7% on financial reasoning at roughly a fourteenth of the cost of a proprietary model, on an internal eval.
• You run high inference volume where an eight-to-one token discount compounds into real money.
• Data privacy rules out sending inputs to Anthropic or OpenAI, and you need weights you control.
• Your task sits inside Inkling's ceiling: instruction following, retrieval-augmented generation, classification, domain reasoning, or multimodal work over text, images, and audio.

Pick a hosted Claude or GPT API when:

• Your work needs frontier coding or hard reasoning the open-weight ceiling does not reach. Choosing a coding assistant instead? Our GitHub Copilot review and Cursor review cover models that outscore a fine-tuned Inkling on SWEBench Verified by 15 to 20 points.
• You lack the labeled data or the ops budget to run and maintain a meaningful fine-tune.
• You need to ship this week and iterate on prompts, not on training cycles.

How to get access

Quickest feel for the model: the Inkling Playground in the Tinker console, free for a limited time, or the base model through OpenRouter at $1 input and $4.05 output per million tokens, per OpenRouter's listing (fetched 2026-07-19). An OpenAI-compatible call looks like this:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/inkling",
    "messages": [{"role": "user", "content": "Summarize this contract clause."}]
  }'

To fine-tune, install the Tinker SDK and set your key, per the Tinker quickstart:

uv pip install tinker
export TINKER_API_KEY="<YOUR_API_KEY>"

Then point a LoRA training client at the Inkling base model:

import tinker

service_client = tinker.ServiceClient()
training_client = service_client.create_lora_training_client(
    base_model="thinkingmachines/Inkling", rank=32
)

Self-hosting the full weights is the part to be honest about. The Hugging Face release ships an NVFP4 checkpoint built for NVIDIA Blackwell systems, per the launch post, so running the full 975B model is not practical on standard cloud GPUs today. Inkling-Small, at 12B active, will be far more accessible once its full weights ship, which Thinking Machines says is still pending. Until then, Tinker or the third-party API partners are the realistic route for most teams.

The bottom line

Inkling is not the model you reach for when you want the best performance available today, and Thinking Machines says as much in its own launch post. It is the one you reach for when you want to own the weights, tune the behavior on data only you have, and serve at scale without paying frontier API rates forever. For a solo dev or a team without a specific fine-tuning use case, a hosted Claude Fable 5 or GPT-5.6 Sol plan is cheaper to run and better this quarter, full stop. For a compliance-bound team sitting on proprietary data and real inference volume, a fine-tuned Inkling on Tinker is the cheaper long-run build, as long as your task lives inside its ceiling. We are holding the score at a provisional 3.5 of 5 and will firm it up once Inkling-Small ships full weights and real fine-tuning reports come in from teams running it in production.