Unlocking Ripple-65B Preview: The 65B Open-Weight Coder You Can Download Today — While Inkling Still Waits

On November 4, 2026, a chip-design startup most people had never heard of dropped a 65-billion-parameter model that, on its own numbers, outscores open-weight rivals up to fifteen times its size. It is Ripple-65B Preview, from Ricursive Intelligence, and the headline is 80.52 on Terminal-Bench 2.1 — agents doing real work in a Linux terminal.

What matters more than the score: unlike Monday's story, you can download this one right now. The weights are on Hugging Face under Apache 2.0 — no waitlist, no "next week," no platform exclusive. That makes it the month's most interesting open-weight release, and the natural follow-up to Inkling 2 Small, the smarter-on-paper model still locked behind a wait.

Note: some coverage dates this release November 6, but the launch post and Hugging Face card went live November 4, 2026 — and every benchmark here is vendor-reported, unreplicated.

What Ripple-65B Preview Actually Is

Strip away the hype and the specs are refreshingly old-school. Ripple is a dense Transformer: 65 billion parameters, all active on every token — no mixture-of-experts routing, no sparse tricks. It trained from scratch on 12 trillion tokens, mostly English prose and code, with 80 layers and a 5,120 model dimension, and supports up to 1,048,576 tokens of context.

It ships as three checkpoints — instruct BF16, experimental NVFP4 for local hardware like NVIDIA's DGX Spark, and a pretrain-only BF16 base model — under Apache 2.0: commercial use, fine-tuning, and redistribution, no royalty, no permission slip.

Ricursive is not a chatbot company. Its real business is AI that designs chips — the team includes former Google engineers from AlphaChip, the reinforcement-learning system that learned to lay out chip floorplans. Ripple is a general-purpose variant of the model behind that platform, retuned for agentic coding. The company says it built Ripple for $32 million end to end, roughly one-tenth the cost of the previous open-weight state of the art on agentic coding. Treat both numbers with caution: the $32 million is Ricursive's own accounting, and the one-tenth comparison is built from Reflection's reported compute commitments, not matched disclosures.

The Benchmarks: Two Wins, and the Fine Print Nobody Reads

Here are the headline numbers, as Ricursive reports them — run at full BF16 precision and maximum effort.

Terminal-Bench 2.1: 80.52 — ahead of Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines' Inkling (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4) — a 0.4-point lead over Beam, which is nearly eight times Ripple's size.
SWE-bench Pro v2.0 Hard: 79.1 — first again, ahead of Beam (76.5) and Muse Spark Open (72.4), on a suite of hard real-world engineering tasks new in September 2026.

Now the fine print. First, no independent evaluation had surfaced — every figure is Ricursive's own, and agent benchmarks swing with the harness. Second, rival scores mix official figures with Ricursive's own reruns, and its own materials disagree: the model card lists SWE-bench Pro v2.0 Hard at 52.3 for Muse Spark Open and 2.0 for Inkling, while the launch graphic shows 72.4 and 39.6. Third, on the official Terminus 2 harness at 128K, most rivals dropped — Beam to 70.8 — while Ripple stayed at 80.5. That strengthens Ripple's case, but treat the rival rows as directional, not gospel.

Ripple-65B vs Inkling 2 Small: The Downloadable Model vs the Promised One

These two models tell opposite stories about what "open" means in November 2026: one is smaller, cheaper, and in your hands today; the other is bigger, smarter on paper, and still behind a door.

	Ripple-65B Preview	Inkling 2 Small
Total / active parameters	65B dense, all active	594B / 17B MoE
Pretraining tokens	12T	Not disclosed
Context window	1M tokens (weights)	1M on card, 64K served on Tinker
Terminal-Bench	2.1: 80.52 (vendor)	4.0: 13.1 (vendor) — different test, do not compare directly
SWE-bench Pro	v2.0 Hard: 79.1 (vendor)	v2 Hard: 86.3 (vendor) — different version, do not compare directly
Artificial Analysis Index	Not listed yet	31
Input price / 1M tokens	$0.20	$0.30
Output price / 1M tokens	$0.60	$1.00
Weights today	Open — Hugging Face, Apache 2.0	Closed — Tinker only, Apache 2.0 promised "next week"

Two things jump out. First, the benchmark rows cannot be compared head-to-head: Inkling's numbers come from Terminal-Bench 4.0 and SWE-bench Pro v2, Ripple's from Terminal-Bench 2.1 and SWE-bench Pro v2.0 — different tests. Anyone quoting "80.5 vs 13.1" as a head-to-head is confused or selling something. Second, on what you can do this week, there is no contest: Ripple is downloadable and Inkling is not. Full Inkling story here: Unlocking Inkling 2 Small.

The Honest Weaknesses: Where Ripple Loses

Ricursive's own table says this is no clean sweep. Outside terminal work, the picture turns mixed fast.

SWE-bench Verified: 76.5% — behind Beam at 81.6% and Muse Spark Open at 78.7%.
DeepSWE v1.1: 44.3% — well behind Muse Spark Open at 60.0%.
SciCode: 39.4% — behind Muse Spark Open at 58.5%.
SWE-bench Pro v1.0: 61.4% — behind Beam at 66.1%.

The "Preview" label is the company's own warning: post-training was light, and the instruct model is a first cut. Reviewers put it plainly — arguably the best US-made open model you can run on one box, but not the best open coding model overall. On SWE-bench Verified and scientific coding, Beam and Muse Spark Open still lead on Ricursive's own numbers. And Kimi K3.1 Pro still leads SWE-bench Pro v2 Hard in Thinking Machines' own table at 88.2% — see Kimi K3.1 Pro vs Claude Opus 5.5.

The Price Is the Real Story

Honestly, the most underrated number here is not a benchmark — it is $0.20 and $0.60, what Together AI charges per million input and output tokens, cached input just $0.02. OpenRouter lists the same. Against Inkling's $0.30/$1.00 on Tinker, Ripple is a third cheaper on input and 40% cheaper on output. Against Kimi's $2.00/$10.00, it is not close.

But there is a catch. OpenRouter currently lists Ripple with a 16K context limit — a fraction of the million-token window in the weights. If your agent workflow needs the full 1M context, the API will not give it to you today; self-host or wait for providers to raise their limits. Together AI serves it as ricursive-ai/Ripple-65B-Preview; Fireworks support was expected soon after launch, not live on day one.

How to Actually Use It Today: A 7-Step Playbook

Download the weights from Hugging Face — instruct BF16 by default, NVFP4 if you run local hardware like a DGX Spark.
Check your VRAM budget first: BF16 needs roughly 130 GB; NVFP4 is the realistic path for a single high-end box.
Start with terminal agent work, not chat — shell tasks, repo chores, multi-step fixes where it can use tools.
Use Together AI or OpenRouter instead of self-hosting at $0.20/$0.60 per million tokens — but remember OpenRouter's 16K context cap today.
Verify before trusting a benchmark: run your own tasks and compare against your current model on your own repo.
Fine-tune freely — that is the point of Apache 2.0. No platform gate: distill it, adapt it to your codebase, ship it commercially.
Watch for independent numbers — Ripple is not on Artificial Analysis yet. Re-check the headline scores before betting a production workflow on them.

One caution: run it with tight permissions. A model this good at terminal work can do real damage with a wrong command — see the real cost of GPT-6 Cyber.

Who Should Care — and Who Should Wait

Download it this week if you want an open coding agent you can run, fine-tune, and ship today — especially if Kimi's $10 output tokens price you out or you are tired of waiting on Inkling's weights. Try the API first if you want to test the terminal-coding claims before buying hardware; $0.20 per million input tokens makes experimenting nearly free. Wait if you need the full 1M context via API, or your work is DeepSWE-style scientific coding, where it trails badly. Skip it if you already run DeepSeek V4.1 Flash or Kimi K3.1 Pro in production and they win on your tasks — a vendor's table is not a reason to migrate.

Final Takeaway: The Open Model You Can Touch Beats the Smarter One You Cannot

Ripple-65B Preview is not the best open-weight model of 2026. It loses to Beam on SWE-bench Verified and to Muse Spark Open on DeepSWE and SciCode, and its two headline wins are vendor-reported and unreplicated, in a table whose rival rows disagree with each other.

But it is the open-weight model that understands what "open" is for: 65B dense parameters, 12T tokens, a million-token window, Apache 2.0, downloadable today, runnable on one box in NVFP4, cheaper per token than the model it chases. Inkling 2 Small may be smarter on paper — 31 on the Artificial Analysis Index, 86.3% on its version of SWE-bench Pro — but paper is all it is until the weights land.

If Inkling's Apache 2.0 release arrives next week as promised, this comparison gets interesting fast. Until then, the model you can actually download wins by default. That is logistics — and in open source, logistics is the whole game.