Inkling 975B: Thinking Machines' Open-Weight Speech Model, Explained

Inkling 975B is Thinking Machines' new open-weight speech-to-text model, weighing in at 975 billion parameters and priced at $6.60 per 1,000 audio minutes on the hosted Tinker platform. Can you run it locally? Not on any consumer GPU sold in 2026 — even at aggressive 4-bit quantization the weights alone dwarf a 12 GB card. The realistic home path is a whisper-family model on hardware like the MSI GeForce RTX 3060 Ventus 3X 12G, and to treat Inkling 975B as an API you call, not a checkpoint you host.

Why an open-weight 975B speech model matters if you already self-host

If you already run a chat model at home, the mental model for speech has been embarrassingly thin. For years the local transcription conversation started and ended with OpenAI's Whisper family — whisper large-v3 at 1.5B parameters, distil-whisper at roughly 756M — and everything else was a hosted API you paid per minute. What Thinking Machines shipped is different in kind: an open-weight audio model at a scale nobody previously bothered to publish for speech, plus a hosted deployment on the Tinker platform whose per-minute pricing is quoted at $6.60 per 1,000 audio minutes. The Artificial Analysis speech-to-text leaderboard now lists the hosted variant alongside the paid incumbents, which is why the search wave hit this week (Artificial Analysis speech-to-text leaderboard).

The question people are actually typing into search is the one this piece answers: yes it is open-weight, no you cannot practically run it at home in 2026, and here is the local transcription rig that makes sense today. That last part matters because if you were already thinking about a self-hosted speech box, the specs of the RTX 3060 12G, an 8-core AM4 CPU and a proper condenser microphone have not changed just because a 975B model appeared. What has changed is that the ceiling of what a hosted API can do just moved up, which sharpens the choice between "call the API" and "run a smaller model locally."

Key takeaways

• Parameters: 975B total, per Thinking Machines' announcement. Compare with whisper large-v3 at 1.5B and distil-whisper at ~756M.
• Hosted price: $6.60 per 1,000 audio minutes on the Tinker platform (as of 2026, per Thinking Machines' quoted pricing).
• Local reality: a 975B model at 4-bit still needs on the order of ~490 GB just for weights — no single consumer card in 2026 fits it.
• What actually fits on 12 GB: whisper-family and distilled variants, plus community fine-tunes.
• Leaderboard placement: listed on the Artificial Analysis speech-to-text leaderboard alongside the paid transcription APIs.

What did Thinking Machines actually release, and under what terms?

Two things shipped in the same window: the model weights themselves, and a hosted endpoint on Thinking Machines' Tinker platform. "Open weights" here means the checkpoint is downloadable from a public host, in this case Hugging Face — the same distribution model most modern open releases use (Hugging Face blog). It does not automatically mean Apache 2.0 or MIT. Several recent open-weight releases carry acceptable-use clauses, attribution requirements, or revenue-threshold gates, so read the model card's licence text before you ship a commercial product on top of it. As of publication in 2026, community reports indicate the licence permits research use and non-commercial deployment, with a separate commercial track routing back through Tinker.

The hosted endpoint is where most people will actually meet the model. Pricing is quoted at $6.60 per 1,000 audio minutes — roughly $0.0066 per minute, or about 40 cents per hour of audio. That is meaningfully above the cheapest hosted whisper endpoints but in the same order of magnitude as the higher-tier transcription APIs, and it is the price you compare against when you sit down to work out whether a local rig makes sense.

How does Inkling rank against hosted transcription APIs?

The Artificial Analysis speech-to-text leaderboard tracks a rotating set of hosted transcription models on word error rate, throughput and price. The exact ranking shifts week to week, so read the live board before you commit; the values below reflect the shape of the market rather than fixed numbers.

The two numbers to fix in your head are the $6.60 hosted price and the fact that whisper-class open-weight models remain the actual local option in 2026. Everything else is a matter of where your workload falls on the accuracy vs. cost curve.

Spec delta: Inkling 975B vs the local whisper family

Two rows do the work here. The fp16 line shows why a single node cannot hold Inkling 975B — that number is bigger than the memory in a stack of H100s. The 4-bit line shows why quantization alone does not save you: even at 4 bits, weights alone need roughly half a terabyte before you count activations, KV cache or audio encoder state.

Can you run it on consumer hardware? Quantization matrix

The honest table is the one that admits when there is no public measurement. All figures below refer to the RTX 3060 12G class; "unverified" means no public benchmark has been posted at the time of writing.

The takeaway is not that quantization is broken — for text LLMs 4-bit routinely brings 70B-class models onto consumer cards. The problem is that 975B is roughly an order of magnitude past what a single 12 GB card can hold at any published quantization. Multi-GPU sharding across a home cluster does not save you either, because you would need on the order of 40 twelve-gig cards just to hold q4 weights, before any activation memory.

Prefill vs generation: why audio encoders behave differently on 12 GB

Text LLMs have a well-known asymmetry — a big one-shot prefill of the prompt, then autoregressive generation that is mostly memory-bandwidth limited. Speech models add an encoder pass on the raw audio, and that encoder is compute-heavy in a way text prefill is not. On a 12 GB card such as the RTX 3060 12G, the practical bottleneck for whisper-family models is rarely VRAM; it is the encoder pass on long audio and the CPU-side audio pre-processing pipeline. That is why bumping to a faster CPU and NVMe scratch space often does more for throughput than swapping the GPU, once the model already fits.

The realistic local path today

If you strip everything down to what fits on a card you can actually buy in 2026, the picture is: pick a whisper-family or distilled model, put it on a 12 GB GPU, and spend the rest of the budget on the input chain and storage. The 12 GB RTX 3060 remains a sweet spot because it clears the VRAM floor for large-v3 with room for a small decoder LM alongside, has enough compute to keep up with real-time transcription for single-speaker audio, and costs a fraction of the current-generation cards. The MSI GeForce RTX 3060 Ventus 3X 12G specifically pairs a 12 GB GDDR6 buffer with a 170 W TGP (TechPowerUp — GeForce RTX 3060 12 GB), meaning it runs on a modest 550-650 W PSU and does not need a boutique case for airflow.

Building the box around it

A local transcription rig is a small computer with disciplined I/O. The GPU is only one of five parts that matter, and skimping on the other four is where most home builds actually lose accuracy or throughput.

• CPU: an AMD Ryzen 7 5800X — 8 cores and 16 threads at 105 W on socket AM4, per the AMD spec sheet — is enough headroom to handle the audio pre-processing pipeline (resampling, VAD, chunking) without stealing cycles from the GPU. Anything with 8 modern cores in the same TDP window works; the point is that a 4-core budget chip will bottleneck long-audio jobs.
• Model scratch space: a WD_BLACK 250GB SN770 NVMe SSD as a dedicated model + intermediate-file scratch keeps checkpoint loads and audio chunk I/O off the OS drive. 250 GB is enough for several whisper variants plus a working directory.
• Bulk audio storage: a Crucial BX500 1TB SATA SSD for raw and archived audio. SATA is fine here — transcription pipelines are compute-bound, not sequential-throughput-bound, and the price-per-terabyte on BX500-class drives is hard to beat.
• Microphone: a HyperX QuadCast 2 USB Microphone placed within about 20 cm of the speaker, in cardioid mode, recording at 16 kHz mono or better. Model accuracy degrades faster with room reverb and clipping than with any bitrate change, so the mic and its placement are the biggest single accuracy lever you have below the model itself.

Perf-per-dollar: hosted $6.60 vs. amortized local rig

At $6.60 per 1,000 audio minutes, hosted Inkling costs about $0.40 per hour of audio. A local rig built around the parts above lands somewhere near $900-$1,100 as a full new build in 2026, or well under that if you already own the case, PSU and storage. The local marginal cost per hour of audio is roughly the electricity — call it 200 W under load for a couple of minutes per hour of audio, or well under a cent per hour.

The crossover is entirely a volume question. Below about a few thousand audio minutes a month, hosted is cheaper once you value your own time to babysit a local pipeline. Above that, and especially if you are already going to run a chat model on the same box, the local path pays back inside a year. The mistake is comparing a fresh $1,000 rig to a $6.60 API bill on month one — the honest denominator is total hours of audio over the useful life of the card.

Worked scenario 1 — Weekly podcast host, ~4 hours of audio a week

You publish one podcast a week, roughly one hour of finished audio and three hours of raw takes and interviews. That is about 15 hours per month, or 900 audio minutes — call it $6 on hosted Inkling. Building a local rig to save $6 a month is obviously silly on the transcription line alone. The right call is to keep hosted, and only revisit if you add video transcription, translation passes or bulk archive work.

Worked scenario 2 — Small research lab, ~200 hours of audio a month

Field recordings, interviews, focus groups — 12,000 audio minutes a month lands you around $80 monthly on hosted Inkling, before any reruns. A local whisper-large-v3 pipeline on an RTX 3060 12G, with a Ryzen 7 5800X handling the encoding and the SN770 as scratch, pays back inside a year even at conservative accuracy parity. You lose the 975B model's specific gains, but for clean single-speaker research audio the accuracy delta to a well-tuned whisper pipeline is small.

Worked scenario 3 — Live captioning experiment

You want sub-second latency captions on a livestream. Neither hosted Inkling (batch-oriented as of 2026) nor a heavyweight whisper large-v3 on a 12 GB card is a great fit. Distil-whisper on the same 3060 12G is the correct target — it trades a small accuracy hit for real-time throughput, leaves VRAM headroom for a decoder LM, and keeps the whole pipeline on hardware you already have.

Common pitfalls

• Assuming quantization saves you. For text LLMs, 4-bit lets you punch above your weight class. For a 975B model, 4-bit still needs roughly half a terabyte of VRAM. Quant is not magic.
• Buying the biggest GPU before fixing the mic. A $1,500 card cannot rescue a $20 headset in a reverberant room. Audio input quality is the single largest accuracy lever below the model itself.
• Underspecing the CPU. Whisper-family pipelines lean on the CPU for resampling, VAD and chunking. A 4-core budget chip will bottleneck long-audio jobs even with plenty of GPU headroom.
• Skipping scratch space discipline. Running model loads and audio chunk I/O on the same drive as the OS turns into pauses and thermal throttling on cheaper SSDs. A dedicated NVMe scratch is a $30-$50 upgrade with an outsized quality-of-life payoff.
• Confusing open weights with open licence. Downloadable does not mean commercially usable. Read the model card licence before you ship.

When NOT to build this rig

Skip the local build if any of these are true. If your audio is a handful of hours a month, the hosted API is cheaper than your time. If you need production-grade streaming captions today, hosted transcription services with dedicated streaming endpoints are still the shortest path. If your workload is dominated by non-English speech in languages where distil-whisper is weak, you may need a hosted model with better multilingual coverage than the current open-weight tier. And if you do not already run other GPU workloads, a card that idles most of the day is hard to justify on transcription alone.

Verdict matrix

• Get hosted Inkling if: you need the specific accuracy of the 975B model, your volume is under a few thousand minutes a month, and you can tolerate batch-mode latency.
• Self-host whisper-family if: you have steady volume (thousands of minutes a month), a 12 GB-class GPU already in the house, and mostly clean single-speaker audio.
• Wait if: you are hoping for a smaller distilled Inkling. It is the obvious next release; a 30-100B distilled variant would collapse the whole "can I run this at home" question the way distil-whisper did for large-v3.

Bottom line

Inkling 975B is a landmark for open-weight speech — the first time an audio model at this scale has shipped as a downloadable checkpoint with a working hosted deployment — but it is not a model you run on your desk in 2026. For hosted use, $6.60 per 1,000 audio minutes is a fair price for the accuracy tier, and the Artificial Analysis leaderboard is the right place to keep an eye on how it moves against paid incumbents. For a local transcription rig, the answer has not changed: put a whisper-family model on a 12 GB card like the RTX 3060, spend on the mic and the CPU, and treat hosted Inkling as an API you call for the jobs your local model cannot handle. That mixed setup — local baseline, hosted escalation — is the shape of practical speech AI at home right now.

Citations and sources

• https://artificialanalysis.ai/speech-to-text
• https://www.techpowerup.com/gpu-specs/geforce-rtx-3060-12-gb.c3682
• https://huggingface.co/blog

This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.