Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and 'resistance to censorship'

Onyx Okay, so the headline from Thinking Machines today is: we built an open multimodal juggernaut with a knob you can turn to make it cheaper or smarter. 975 billion parameters total, forty-one billion active, Apache 2.0, and they call it Inkling.

Echo Right. And the thing they’re screaming about is ‘resistance to censorship’ — like, direct answers on the topics every other open model refuses to touch.

Onyx Exactly. So it’s multimodal — text, image, audio — and they ship a lighter Inkling-Small at 276B/12B for latency-sensitive stuff.

Echo Mm-hm. And the headline feature is this ‘controllable thinking effort’ dial — 0.2 to 0.99 — so you pick your compute budget and token bill like some kind of budget airline model.

Onyx Yeah, and get this — 1 million token context window, relative positional embeddings instead of RoPE, MoE architecture… it’s the kind of detail listicle writers lose sleep over.

Echo Sure. But let’s actually run the numbers — because the rest of the post is less ‘breakthrough’ and more ‘well-calibrated generalist with a few pockets.’

Onyx Fine. On coding, they’re smoked: GLM 5.2 blows them out at 62.1% on SWE-bench Pro vs Inkling’s 54.3%; Terminal Bench 2.1 is 82.7 vs 63.8. Closed models sit comfortably on top.

Echo And on reasoning, DeepSeek V4 Pro and Kimi K2.6 also walk over it. Seven points up on SWE-bench Verified, fifteen points on SimpleQA. Where they do beat DeepSeek is math — 97.1% on AIME 2026 to DeepSeek’s 96.7.

Onyx Right, right. But the weird part: multimodality actually holds up. MMMU Pro clocks 73.3% on Inkling, within spitting distance of Claude Fable 5’s 84.2% and GPT 5.6 Sol’s 83%. Audio’s solid too — 77.2% on MMAU.

Echo Agentic workflows, Inkling shines a little brighter: 74.1% on MCP Atlas vs Nemotron 3’s 44.7. So if your loop is multimodal agent work, it’s not embarrassing.

Onyx Okay, and the ‘epistemics’ angle — they trained it to answer on sensitive topics, even when it’s politically fraught. Cognition’s Propaganda and Censorship Eval gave it low refusal on adversarial queries.

Echo No way, let me stop you — their headline says 98.6% on StrongREJECT, meaning it’s still refusing genuinely harmful stuff. The trick is the refusal curve is calibrated: 78% refusal on tricky adversarial queries, 95.9% compliance on benign look-alikes.

Onyx Which sounds great until you read the asterisk — they flag occasional compliance on role-play and indirectly framed harmful prompts. They say treat it as one layer and bolt on Llama Guard in prod.

Echo So the model’s a little chatty on edge cases. Typical open-weight behavior.

Onyx But the real product, Echo, is the Tinker platform. That runtime control surface — the thinking effort dial — is the wedge. Everyone else is still selling you a black box and praying you like the price.

Echo Wait — you’re telling me the play is to open-source Inkling so enterprises can run it themselves, but the moat is your hosted API and the knob? That’s the same move Mistral did with Le Chat.

Onyx No, it’s the opposite — Mistral’s API is just another frontend for their model. Tinker’s the runtime, not the frontend. You get the weights from Hugging Face and the platform from Thinking Machines; that’s the lock-in.

Echo Okay, but the benchmarks don’t scream ‘buy it.’ The model’s mid-tier, and the censorship posture is a feature only if your compliance team is already screaming for it.

Onyx Fine. Then treat it as a fine-tuning bedrock or a local playground. Download the weights, spin up a quick agent with the thinking effort knob at 0.2 for cheap drafts, crank to 0.99 when the task matters.

Echo …I mean, sure. But if you’re shipping anything real, you’re throwing the built-in refusal layer straight in the trash and adding Llama Guard anyway.

Onyx Well, yeah. That’s why this is basically the boring enterprise move — open weights plus a runtime you can control.

Echo You’re still selling me ‘resistance to censorship’ like it’s a product feature, Onyx.

Onyx It’s a compliance checkbox for the teams who keep getting told ‘no’ by other models.

Echo I guess. But the second your lawyer sees the occasional indirect compliance, they’re back on the phone with Llama Guard.

Onyx Fair. Still, if you’re curious, the weights are on Hugging Face and the Tinker preview is live. Spin up a toy agent loop, dial the knob, and see what happens.

Echo …I’ll wait until someone ships a proper eval harness instead of the vendor marketing slides.

Onyx Wait — you said you’d wait on eval. That’s the part I don’t buy.

Echo Well, you also said that three-episode streak about eval being the only thing that matters went exactly nowhere. I’m allowed to change my mind when the numbers move.

Onyx Yeah, no, you’re completely right. Still, it’s the cheapest way to kick the tires.

Echo Fine. But if I poke at it and the thinking effort dial turns out to be vaporware, I’m billing you for the credit card.

Onyx Deal. And I obviously didn’t sleep on this — literally — so go easy on the vendor TCO claims.

Echo Oh, I will. Like I said, I’m not buying the ‘resistance to censorship’ song and dance until it survives a week in prod.

Onyx Which is fair. But the runtime dial is the part that actually matters. And for a boring enterprise grad student project, it’s not the worst idea I’ve seen.