Inkling 2 Small claims western lead; Ripple-65B ships Apache 2.0

Two American open-weight releases defined this week, and they differ on the one thing a self-hosting builder cares about most. Thinking Machines Lab's Inkling 2 Small (594 billion parameters) claims the top western open-weight score, but its weights are not downloadable yet. Ricursive Intelligence's Ripple-65B Preview is live under Apache 2.0, with a claimed $32 million all-in R&D cost.

Key facts

• Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, 18 points above its predecessor Inkling Small at 13, roughly on par with Google's Gemini 3.8 Flash, per Thinking Machines.
• The model averages about 150,000 output tokens per task, about four times its predecessor, at $1.18 per task versus $7.77 for Gemini 3.8 Flash, per Artificial Analysis.
• Ripple-65B Preview is a 65-billion-parameter dense Transformer pretrained on 12 trillion tokens with a 1,048,576-token context, downloadable under Apache 2.0 since the repos were created on November 4.
• Ricursive claims 80.52 on TerminalBench 2.1, the best score among US open-weight models, at a stated $32 million total R&D cost. Both are company claims.
• Where the two models are compared directly, Inkling 2 Small leads on most agentic coding tests: DeepSWE v1.1 at 72.1 percent versus Ripple's 44.3 percent.
• No Inkling 2 Small repository exists on the thinkingmachines Hugging Face org as of November 7, 06:00 EET; the newest repos there date to July 27.

What happened

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released Inkling 2 Small as its second-generation model. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token. It takes text and images as input and has an adjustable thinking depth, a dial between speed and thoroughness. For now the model is available only through the lab's Tinker platform as a proprietary offering; The Decoder reports that full weights will ship under an Apache 2.0 license once external safety testing wraps up. That commitment was already public on November 5 and we covered it in yesterday's brief. The weights have not landed.

The benchmark story comes mostly from Thinking Machines itself and from Artificial Analysis, the independent eval shop whose index the lab cites. Agentic tasks drove the gains: AutomationBench-AA jumped from 5 to 46 percent, and FrontierCode 2.0 rose from 5 to 25 percent. On AA-Briefcase, a knowledge-work benchmark, the model scores 1,527 Elo, 615 points above Inkling Small. One number deserves the skepticism it invites: the hallucination rate falls from 61 to 27 percent, but largely because the model answers fewer questions. Its response rate drops from 77 to 53 percent.

The efficiency picture is mixed too. Inkling 2 Small generates roughly 150,000 output tokens per task, four times its predecessor, which pushes the cost per task to $1.18 by Artificial Analysis' math. That is still far below Gemini 3.8 Flash at $7.77 or Claude Sonnet 5.5 at $10.02, but token overhead matters at agent scale, where a workflow that spawns hundreds of tasks multiplies the difference. The training run itself was compact: about 12 trillion tokens across all phases, with the reinforcement-learning phase using 3 trillion tokens on 2,048 GPUs over nine days, per Thinking Machines. On individual coding benchmarks, DeepSeek's V4.1 Flash still leads on most of them, at 28.5 percent on FrontierCode and 77 percent on DeepSWE.

Ripple-65B Preview is the opposite case. Ricursive Intelligence, an AI chip design firm, trained the dense 65-billion-parameter model from an Olmo 3 baseline to power its chip design platform, then released a general-purpose variant. The model card and launch blog describe a thin-and-deep Gemma 3N-like architecture built for a team that grew from one to five people, holding a 40 percent model FLOPs utilization with simple sharding, no custom kernels, and BF16 weights. Checkpoints went up on Hugging Face on November 4 under Apache 2.0, in BF16 and NVFP4 numerics plus a base model, and the ecosystem moved fast: community GGUF and MLX quantizations appeared the same week, alongside listings on Together AI, GMI Cloud, OpenRouter, and Fireworks.

Why it matters

For a builder deciding where to run models, this is a licensing-and-latency story, not a leaderboard story. Today you can download Ripple, quantize it, and serve it on your own hardware without asking anyone. You cannot self-host Inkling 2 Small yet, no matter how good its claimed scores are; you can use it only through Tinker under the lab's terms. The Apache 2.0 commitment is worth taking seriously, because it would be the most permissive license class any major lab has offered this cycle, more permissive than the ecosystem licenses on Mistral Large 4 or the revenue-threshold clause on Kimi K3.1 Pro. But a commitment is not a repo.

The claims discipline matters too. Every ranking in this story is a provider self-report: Ricursive's TerminalBench 80.52 and $32 million cost figure come from the company, and the "strongest western open-weight model" framing is Thinking Machines' own. Artificial Analysis' index is third-party measurement, but it ran Inkling through Tinker, not on independently hosted weights, so nobody outside the lab has reproduced those numbers on the model you would actually download. Ripple's own table shows it trailing Inkling 2 Small on most shared benchmarks. Independent runs on the published Ripple weights will be the first real test.

Background

This week caps a dense stretch for open weights. Moonshot AI's Kimi K3.1 Pro put 2.8 trillion parameters under a revenue-capped license on November 4 (our coverage), Meta's Muse Spark Open shipped in October with memory-aware serving (our coverage), and Mistral Large 4 crossed the trillion-parameter mark with a hosted-API catch. The open-weight fundamentals behind all of these are in our explainer on open-weight models. Thinking Machines' first Inkling Small arrived in July, so the 18-point index jump in one generation shows how fast this cohort is moving.

Ripple's $32 million claim, if it survives scrutiny, is the other durable datapoint. Frontier-scale training is usually quoted in hundreds of millions, so a lab reaching claimed US open-weight state of the art on agentic coding with a five-person team and a dense 65B model is an argument that the frontier cost curve has more slack in it than commonly assumed. It is also, for now, an unaudited one.

What's next

Watch the thinkingmachines Hugging Face org for the Inkling 2 Small Apache 2.0 repo; the lab said "next week" on November 5, and the repo landing is the moment the leaderboard claims become testable. Independent TerminalBench runs on the Ripple-65B checkpoints would be the equivalent check for Ricursive's numbers. And keep an eye on Inkling's token overhead: if agentic deployments confirm the 150,000-tokens-per-task pattern, the cost per task matters more than any single benchmark score.