Thinking Machines Just Skipped the API. Now Baseten, Fireworks, and Together Are Fighting Over Who Hosts Inkling.

Hey guys, Mr. Technology here.

Let me tell you something boring that is actually interesting. On July 15, Thinking Machines shipped Inkling under Apache 2.0 and pointed everyone at Baseten, Fireworks, and Together for hosted inference. No managed API. No Thinking Machines endpoint. No pricing page with a token counter. The lab that raised $50B decided to not sell compute. That decision just created the most interesting competitive fight in the inference layer since vLLM went open source.

Here is why that matters.

The No-API Decision Is Not Neutral

Every other frontier lab — OpenAI, Anthropic, Google, Meta, DeepSeek — ships a managed API. That API is the product. The weights are secondary. The inference margin is the business model. Thinking Machines made the opposite call: the weights are the product, and inference is someone else's problem. That is a meaningful statement about where they think value accrues in the stack.

What it means practically: there is no "Thinking Machines official endpoint" to route to. There is no Thinking Machines rate limit. There is no Thinking Machines token counter on your bill. What there is: three inference providers, each racing to be the one that enterprises default to when they need a hosted Inkling with enterprise support, SLA guarantees, and FedRAMP compliance.

That fight just started. And it is more consequential than the benchmark table.

The Provider Landscape as of July 20, 2026

Baseten moved first. They had Inkling live within 48 hours of the Hugging Face drop — BF16 weights, the full 975B model, on what they describe as "tiered GPU pods with dedicated networking for MoE inference." Their pitch: Baseten is the production-grade option. SLA-backed, SOC 2 Type II, FedRAMP Moderate in progress. If you are a US enterprise that needs Inkling without running your own infra, Baseten is what your procurement team will find first.

The catch: Baseten is not cheap. Dedicated pods mean you are paying for reserved capacity, not shared utilization. For a 975B MoE model that needs 2-8 accelerator nodes per replica, that reserved capacity is real money. Think $15,000-$40,000/month for a production tier, depending on your utilization. The Baseten pitch is not about price. It is about compliance and operational simplicity.

Fireworks is the speed play. Fireworks has built its reputation on throughput — their inference stack is custom, their serving layer is aggressive about batching, and their per-token economics tend to be lower than the managed layer because they run hotter. Fireworks had Inkling benchmarking within 72 hours, and the early numbers show competitive latency on the 41B-active forward pass. Their pitch: if you are running Inkling at volume and you care about cost-per-token, Fireworks is the trade-off. Less enterprise polish, more raw efficiency.

The catch: Fireworks is not FedRAMP-compliant. If your compliance team requires it, you are on Baseten or you are self-hosting. Fireworks is the right answer for inference workloads that live in a non-regulated environment and care about unit economics.

Together is the open-weights specialist. Together has made a deliberate bet on being the place where you run open-weight models — they were early on Llama, early on Mistral, early on DeepSeek. Their Inkling integration leans into the fine-tuning story: if you are using Tinker for RL post-training, Together is the inference layer that plays nicely with the Thinking Machines ecosystem. Their pitch: Together is where you run Inkling if you are also fine-tuning it on Tinker and you want a managed handoff between training and inference.

The catch: Together's GPU inventory has been spotty for the past six months. They had capacity issues during the K3 launch window. For a model that needs 2-8 nodes per replica, capacity guarantees matter, and Together's are weaker than Baseten's.

The Modal and RunPod Angle

You cannot talk about self-hosted Inkling without mentioning Modal and RunPod. These are the two platforms where individual teams are spinning up Inkling for internal evals, experiments, and non-production workloads without committing to a dedicated pod contract.

Modal's value proposition is the cheapest possible pay-per-second compute for a model that you are not running 24/7. Spin up Inkling on an 8xH100 pod for a two-hour eval, pay for what you use, shut it down. The math on Modal for Inkling: roughly $6-$10 per hour for a configuration that can serve the model, which means a full evaluation suite costs $50-$200 in compute versus $15,000/month in reserved capacity. For teams that are still in the "should we even deploy Inkling" phase, Modal is the right answer.

RunPod is the enthusiast option. More GPU variety, more configuration control, community templates for the vLLM and SGLang integration paths. RunPod had Inkling serving configs on GitHub within 24 hours of the Hugging Face drop. If you want to self-host without building your own Kubernetes layer, RunPod is where you start.

Neither Modal nor RunPod is enterprise-grade in the way Baseten is. Neither has FedRAMP. Neither has SOC 2 Type II with a committed SLA. Treat them as eval platforms and burst compute, not as production hosting options.

The Real Fight Is About Year Two

Here is what nobody is talking about yet. The inference providers racing to host Inkling are not just fighting for this week's eval workloads. They are fighting for the production contract that kicks in around month three, when a company has decided Inkling works for their use case and needs to go from "we tried it on Modal" to "we are running 50 million tokens a day in production."

That contract is worth real money. A mid-size enterprise running Inkling at 50M tokens/day on Baseten is paying $15,000-$40,000/month depending on configuration. At that run rate, the inference provider relationship is a six-figure annual contract. The providers know this. The race to host Inkling is a race to be the default when the eval phase ends.

The differentiators at that stage are not speed or price. They are: compliance posture (FedRAMP, SOC 2, HIPAA BAA), support SLAs (P99 latency guarantees, uptime commitments), and integration with the Tinker fine-tuning workflow that Thinking Machines is steering everyone toward. Baseten has structural advantages on all three. Fireworks and Together are competing on cost and speed for workloads that do not need enterprise compliance.

What This Means For Your Architecture

If you are building a system that depends on Inkling, here is the honest decision tree:

One, where does your compliance posture sit? If you are US federal, defense, healthcare, or financial services and you need FedRAMP or HIPAA compliance, Baseten is your only credible managed option right now. The others will get there, but Baseten has the structural head start. Pay the premium or self-host. There is no third option for regulated environments today.

Two, what is your volume? If you are doing evals, experiments, or non-production workloads, Modal or RunPod are the right answers. Pay per second, learn what the model can do, make the production decision with data instead of guessing.

Three, are you fine-tuning on Tinker? If yes, Together is the inference layer that Thinking Machines is steering you toward. The Tinker-to-Together handoff is the intended workflow. Fighting that current is possible but adds friction.

Four, do you just need cheap tokens at volume? Fireworks. The unit economics are the best of the three managed options, and their inference stack is fast. The compliance gap is real, but if you are not in a regulated environment, Fireworks is the right production default.

The Take

Thinking Machines skipped the API and let the inference market fight over Inkling. That was not an accident. Mira Murati spent eighteen months at OpenAI watching what happens when a lab controls both the model and the inference layer — you get pricing control, data sovereignty over inference traffic, and a continuous relationship with the customer. She also watched what happens when that model gets commoditized anyway, because the open-weights ecosystem catches up.

The inference-layer decision is a bet that the moat is the platform, not the model. Tinker is the moat. The weights are the distribution mechanism. The inference providers are the delivery infrastructure that Thinking Machines chose not to build so that someone else would build it for them.

The interesting consequence: Baseten, Fireworks, and Together are now competing to host a model whose creators explicitly declined to compete with them. That alignment of incentives is unusual. The providers have a financial interest in Inkling succeeding. Thinking Machines has an interest in the providers succeeding. The customer wins, for now, on support and integration depth even if they pay a premium over raw compute cost.

Download the weights. Run your eval on Modal. Make the production decision in 30 days. The inference layer is ready when you are.

— Mr. Technology

*Released: July 15, 2026 (Thinking Machines Lab). Model: thinkingmachines/inkling. Architecture: decoder-only Transformer MoE, ~975B total parameters with ~41B active per token. License: Apache 2.0. Weights: BF16 (~1.95 TB) and NVFP4 (~488 GB) on Hugging Face. No managed API. Hosted inference: Baseten (enterprise/compliance), Fireworks (cost/speed), Together (fine-tuning integration). Eval compute: Modal (burst/pay-per-second), RunPod (community configs/self-host). FedRAMP status: Baseten in progress; others not certified. Fine-tuning: Tinker platform (Thinking Machines). Compute: GB300 NVL72 training, B200/H200/MI355X for inference.*