Ripple-65B open coding model claims 1M-token context, but its DGX Spark script caps at 128K

Ricursive Intelligence released Ripple-65B Preview on November 4, a 65-billion-parameter dense coding model the company says it pretrained from scratch on 12 trillion tokens. The weights sit on Hugging Face under an Apache 2.0 license, and the headline claim is an 80.5 score on Terminal-Bench 2.1. The catch, laid out in a dev.to analysis by Gabriel Anhaia, is that the advertised one-million-token context only appears on server hardware: Ricursive's own script for NVIDIA's DGX Spark desktop box stops at 131,072 tokens.

What shipped

Four repositories accompany the release, per the model card: a BF16 instruct model, a BF16 base model, an NVFP4 build labelled experimental that quantizes only the MLP layers while attention stays in BF16, and a DFlash2 draft model used for speculative decoding.

The architecture is new enough that it ships as code rather than a built-in Transformers type. The repo sets its own model_type and points to its own modeling file, so loading the weights with Transformers requires trust_remote_code=True, which executes the publisher's Python on your machine. The dev.to post's advice applies: read that file before running it.

Context claims also vary within Ricursive's own documents. The card's introduction says "up to 1M," while its training table lists a 2M sequence length for the long-context stages.

A scorecard with caveats

The 80.5 Terminal-Bench figure puts Ripple ahead of Reflection's Beam (501B parameters, 80.1), Meta's Muse Spark Open (816B, 76.4) and Thinking Machines' Inkling (975B, 63.8). But as VentureBeat points out, these are self-reported numbers from separate vendors rather than a controlled head-to-head, and Ricursive's table mixes harnesses: Terminus 2 for some benchmarks, Inspect AI for SciCode, and OpenCode where none is named.

Ripple also loses several rows on its own card. Beam leads SWE-bench Verified 81.6 to 76.5 and SWE-bench Pro v1.0 66.1 to 61.42. Muse Spark Open leads DeepSWE 1.1 (60.0 versus 44.3) and MCP Atlas (89.2 versus 71.67). Ripple's SciCode score of 39.4 is the lowest of the five models listed. VentureBeat additionally found a discrepancy between Ricursive's launch graphic and its card over Inkling's SWE-bench Pro v2.0 Hard score: 39.6 in one document, 2.0 in the other.

The size comparison needs a footnote as well. Beam is a mixture-of-experts model activating 23B parameters per token, while Ripple is dense, so every generated token passes through all 65B parameters — on hardware Ricursive itself describes as short on memory bandwidth compared with datacenter GPUs.

Three ways to run it, three context windows

The same weights ship with different limits depending on the route. On OpenRouter, served by GMICloud in BF16, the listing shows the full 1,048,576-token window with 131,072 maximum output, priced at $0.20 per million input tokens, $0.60 per million output and $0.05 for cached input; VentureBeat had earlier reported a 16,384-token limit there, so the listing is worth rechecking before sizing prompts. Ricursive's own server recipe runs SGLang 0.5.19 on two B200 GPUs with the full window and eight concurrent requests per port. The DGX Spark script fixes context at 131,072 tokens and handles one request at a time.

Together AI lists the model as coming soon, and per VentureBeat, Fireworks — which the launch page said would go live on release day — had no confirmed rate.

What local actually means

The Spark script checks specifically for the box's GB10 GPU (compute capability 12.1) and errors out on anything else. It wraps SGLang's engine inside the Python process, with no HTTP server, no port and no streaming, so any coding agent expecting an OpenAI-compatible endpoint would need a server put in front of it. The dev.to author notes he has not run the script himself.

A social announcement also named MacBooks as a target, according to VentureBeat, but the release specifies no laptop memory configuration, speed or usable context. The Hub already hosts nine community quantizations, including MLX and GGUF builds, none of them from Ricursive, and Ricursive's own cards do not mention them.

Why it matters

An Apache 2.0 dense 65B model that competes with models eight to fifteen times its total size on agentic coding benchmarks is a real data point for local AI tooling — the weights, a quantized variant and a speculative-decoding draft model are all available. But the fine print matters as much as the launch. The gap between "1M context" and the 128K local window, benchmarks that mix vendors' self-reported scores and differing harnesses, zero independent Terminal-Bench runs cited so far, and preview or experimental labels on the releases all point the same way: test on your own hardware and your own tickets before building around it.