Thinking Machines Lab, Mira Murati's company, launched Inkling 2 Small on November 5: a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, aimed at agentic coding, vision and long-horizon knowledge work. For now it lives only on Tinker, the lab's sampling and fine-tuning platform, with closed weights. The lab says it intends to publish them under Apache 2.0 "next week," once external safety testing and red-teaming are done. If that happens, an open-weight field that is mostly Chinese in the lab's own comparison table gets a serious American entrant. Until then, this is an API model with a promise attached.
Five hundred ninety-four billion, and still "Small"
The model card fills in the shape: a 48-layer decoder-only transformer with a sparse MoE feed-forward backbone, each token routed to 8 of 384 experts, attention split between local and global layers, and a context window of up to 1M tokens. Images enter through a vision transformer (a 292M-parameter encoder based on Perception Encoder, per one release tracker); output is text only. The card lists MXFP8 and NVFP4 numerics. Seventeen of 594 billion means about 2.9% of the weights work on any given token, sparser than the first Inkling's 41B active of 975B in July, roughly 4.2%.
Per the announcement, the first pretraining phase pushed 8.4T tokens through NVIDIA GB300 NVL72 racks in 68 hours, natively in NVFP4, followed by 3.8T tokens at 512K context and smaller phases of agent and reasoning trajectories. Post-training was one joint RL run: 1,075 steps, more than 17M rollouts, 2,048 GPUs, nine days, over 4×10²³ FLOPs.
Then there is the name. Inkling Small, released in late July, had 276B total and 12B active parameters. Its successor in the "Small" slot is more than twice as large. Model naming was always aspirational; here it is merely relative.
A strong scoreboard, kept by the home team
Every benchmark below comes from the lab's own table on the model card, which sets the model against twelve open and closed rivals, among them GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5. None of it has been independently reproduced.
On DeepSWE v1.1, the coding benchmark the lab leads with, Inkling 2 Small scores 72.1%, third of the ten open-weight models listed: behind DeepSeek-V4.1-Flash (77.0%) and MiniMax-M3.1-Flash (73.7%), ahead of MiMo V2.6 Pro (71.9%), GLM 5.3 (69.0%), Kimi K3 (68.5%), MiMo V2.6 Flash (67.9%), GLM-5.3-Flash (63.4%), Mistral Large 4 (61.7%) and Muse Spark Open (60.0%).
Elsewhere the picture is respectable rather than dominant. On SWE-bench Verified it posts 82.2%, well short of GLM-5.3-Flash (92.0%) and GLM 5.3 (95.4%). On the Artificial Analysis index (v5.0) it scores 30.6, behind Claude Haiku 5.5 (36.5) and DeepSeek-V4.1-Flash (32.6). Its outright wins cluster in vision and agentic tooling: PerceptionBench at 64.7%, Finance Agent v2.1 at 67.5%, and BrowseComp with context management at 90.2%, though on that row only Mistral Large 4 (84.5%) reported a score at all. Winning a race most of the field skipped is still winning, just a quieter kind.
One trap: Inkling Small scored 40 on the Artificial Analysis index in July, yet the announcement says its successor beats both predecessors there while the table lists 30.6. The scale has evidently been reset with v5.0; compare like versions only.
What Apache 2.0 weights would actually buy
Today the model card lists the licence as Tinker Terms of Use, with access through Tinker's API for sampling and fine-tuning plus a browser playground. Serverless pricing, as reported by AI Reiter citing the lab's pricing page, is $0.30 per million input tokens, $0.03 cached and $1.00 output at 256K context. The first Inkling listed at $3.74 input and $9.36 output at the same 256K tier, per Artificial Analysis. If the new figures hold, output costs about a ninth of what it did in July.
Apache 2.0 is the permissive end of the spectrum: commercial use, modification, redistribution, a patent grant, no user thresholds. What it cannot do is shrink the model. The 17B active figure governs compute per token; the 594B total governs memory. At 4 bits the weights alone come to roughly 300 GB before scale factors and KV cache, at 8 bits roughly 600 GB (my arithmetic). That is a multi-GPU node, not a workstation. For the first Inkling, vLLM shipped day-zero NVFP4 and BF16 support and Unsloth's 1-bit quants cut it from 1.9TB to 270GB; for v2, no VRAM guidance or serving recipe exists yet.
Fine-tuning is the real pitch. The announcement claims the model matches Inkling Small's final fine-tuned performance within 75 training steps. With weights you could train adapters on your own hardware; a full fine-tune of 594B parameters stays a cluster-scale job. The weights would be free. The convenient place to train them would still be Tinker. That is the business model, and not a bad one.
The lab, the money, and the asterisks
Murati, OpenAI's former CTO, raised a $2 billion seed at a $12 billion valuation in July 2025. Per the same Runtime Wire report, a fall 2025 push for $50 billion-plus was abandoned, The Information later reported the lab raising more than $1 billion above $40 billion, and Wired reported two co-founders leaving to rejoin OpenAI. Since July the lab has shipped three models, which is one way to answer a valuation question.
The asterisks are plain. The weights are not out, and "next week" from a Thursday announcement means after November 9 at the earliest, conditional on outside testing. The benchmarks are self-reported, with blank cells doing some of the work. The lab's internal misbehavior score fell to 14.9 from 66.8 for Inkling and 75.2 for Inkling Small, but GPT-6.1 Sol scores 0.8 and the lab concedes it still trails closed models on alignment. The model card puts its dangerous capabilities "materially below frontier capabilities," and lists the familiar limits: hallucination, loose instruction-following, decay in long conversations.
The 30,000-foot read
Inkling 2 Small is not the best open model on most rows of its own table, and it does not need to be. It needs to be good enough, cheap enough and permissively licensed enough that teams choose it as the base they fine-tune, and on the evidence so far it plausibly is.
What can be watched is short: whether the weights land next week under the licence promised, whether independent harnesses reproduce the coding numbers, and whether the serving community squeezes a 594B MoE into sane hardware budgets as it did for the first Inkling. Everything else is a press release waiting to be checked.
Sources: Thinking Machines Lab, "Introducing Inkling 2 Small" (launch date, parameters, Tinker-only access, Apache 2.0 plan and safety-testing condition, training phases and RL run, fine-tuning claim, misbehavior scores); Inkling 2 Small model card (architecture, current licence, all benchmark figures and comparison models, safety findings, limitations); Runtime Wire on the launch (closed weights at launch, "next week" plan, $2B seed); Runtime Wire on funding talks ($12B seed valuation citing TechCrunch, abandoned $50B ask, The Information's $40B report, Wired's co-founder report); AI Reiter (Tinker serverless pricing, absence of serving guidance); The Decoder (Inkling Small parameters, July AA index score); Latent Space AINews (first Inkling parameters, Artificial Analysis pricing, vLLM and Unsloth serving notes); AI TLDR release tracker (vision encoder size). All benchmark figures are lab-reported and not independently verified. Tinker pricing is single-sourced to a secondary site citing the lab's pricing page, which I could not open. Funding details are secondary reports relayed by Runtime Wire, not primary filings. The first Inkling's parameter count varies slightly across posts (975B in most); self-hosting memory figures are back-of-envelope arithmetic. Weights had not been released as of the announcement; check before relying on the Apache 2.0 licence.