Ripple-65B Preview: falsify open weights and TerminalBench reruns, not “one-tenth cost” copy

Ricursive Intelligence released Ripple-65B Preview, calling it a general-purpose variant of the proprietary model behind its chip-design platform, with hosting on Hugging Face, Together AI, GMI Cloud, and OpenRouter (Fireworks said to follow the same day). The news core should be downloadable 65B open weights plus a checkable agentic-coding score, not “full-stack self-improvement” copy. The company puts TerminalBench 2.1 at 80.52 as U.S. open-weight SOTA and puts full from-scratch R&D at $32M, about one-tenth its estimate of the prior open-weight SOTA. Both need outside checks.

Hard facts (company post and HF card, accessed 5 Nov 2026): - Scale: 65B dense Transformer; 12T pretrain tokens; claimed up to ~1M context (HF card details per-stage sequence lengths).

• Architecture: thin-and-deep, d_model 5120, 80 layers; dense instead of MoE to cut train/serve complexity; claimed ~40% MFU with simple sharding, no custom kernels, no quantization (BF16). Also SWA/global ratio tweaks, FFN activation clipping, gated attention, global NoPE, value residuals.
• Scores (company table): TerminalBench 2.1 = 80.52 vs Reflection Beam 501B 80.1, Meta Muse Spark Open 816B 76.4, Thinking Machines Inkling 975B 63.8, NVIDIA Nemotron 3 Ultra 550B 56.4.
• Cost: $32M full R&D from scratch (ladder, mixes, evals, synth data, post-train recipe, final hero run). Footnote cites reports that Reflection-class players committed ~$150M/month compute and that Beam took ~4 weeks each for pretrain and post-train.
• RL: no purchased post-train data; ~1.4M rollouts to 80.5, about 70× fewer than Reflection Beam per the company.
• Artifacts: BF16 instruct, experimental NVFP4, BF16 base; HF lists midtrain / long-context / SFT / RL token and step counts.

Evidence grade: primary product/research post plus an HF weight page. Scores and costs are company-reported; no same-day third-party TerminalBench rerun or audited cost ledger. Choosing dense over MoE is framed as an infrastructure bet, paired with MFU and local-device (e.g. DGX Spark) narratives. Mechanically Ricursive stresses a model ladder from Olmo 3, adding scaling rules, pre+post norm, gated attention, global NoPE, value residuals, and upsampled code. It claims up to ~16× better compute efficiency on code-related dolma_stack CE vs the Olmo 3 ladder at 4B scale. That is a small-scale ladder claim; the jump to 65B still needs a publicly checkable hero run. Steelman: if HF weights download cleanly and independent evaluators remake ~80.5 under the same TerminalBench 2.1 protocol, then “a 65B dense open model beats several larger U.S. open models on that benchmark” stands. The “one-tenth cost” line needs audited spend for the comparison object, not footnote media reports of compute commitments.

The HF card adds a stage table: 12T pretrain (~877,500 steps at 8K), midtrain 200B, long context 6.5B, SFT 183B, LC SFT 13.5B, and RL about 700 steps. That gives “relatively light Preview post-training” a numeric scale. If third parties only remake the instruct weights and skip the base, the from-scratch pretrain claim stays only half-tested.

Best steelman: a small team (HF card: grew from 1 to 5 during development) trained a locally deployable 65B from scratch and shipped base / instruct / NVFP4, weakening the default that only giant MoEs compete on agentic coding. Steelman risks: “U.S. open-weight SOTA” depends on the company’s chosen peer list and date cut, without a claim of covering every relevant release. $32M is an internal full-cost figure without third-party audit; the Reflection “~$150M/month” contrast is an estimate, not same-GAAP books. Preview post-training is described as relatively light and coding-focused; non-coding generalization evidence in the post is thin. Effective quality at 1M context vs the nominal length lacks an independent long-context table.

For the open-weight ecosystem, six-month checks are lasting downloads, third-party TerminalBench and related agent evals, and real latency/throughput on Together / OpenRouter-class hosts. For the hardware story, keep any internal chip-design variant’s verifiable design outputs separate from the general Preview score. For buyers, local fit of a 65B dense model is a memory/bandwidth constraint, not a “small means strong” slogan.

For peer labs, the release pressures same-protocol open-weight comparisons: if larger U.S. open models trail a 65B on TerminalBench 2.1, the next move is a stronger public rerun or a clear protocol diff. For users, pick by task success and local VRAM, not parameter rank.

Test downloadable weights and independent TerminalBench reruns first. One-tenth cost and the chip loop are stronger, harder-to-falsify add-ons. If the score holds, this is a weighty open release; if scores or weights fail, it remains a company benchmark post. 【No.2】