What Four U.S. Coding-Model Releases Disclose: Licenses and Weight Access Differ, and Each Release's Benchmark Table Is the Vendor's Own

Background

Four model releases aimed at coding and agent work were reported between October 5 and November 6, 2026, and each source places its model in a U.S. context. TechCrunch wrote on October 5 that Reflection AI, a Brooklyn-based startup, was unveiling Beam, a 501-billion-parameter model with 23 billion active parameters and a 1 million token context window, and that Reflection said it would release the weights that month. SiliconANGLE wrote on October 15 that Meta had released open weights for Muse Spark, which had been proprietary since its April debut, as Muse Spark Open, and that Meta describes it as the most capable open-weight model to come out of a U.S. lab so far. VentureBeat wrote on November 4 that Ricursive Intelligence’s Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks. The New Stack, in an article dated November 6, wrote that Cognition’s SWE-2 Flash, which it says was released on a Wednesday (November 4 by the Herald’s calendar count), is built on Thinking Machines Lab’s open-weight Inkling-Small while Cognition keeps its own weights private and describes the model as fully American.

The Herald has covered related releases and licenses: Ripple-65B Preview, Moonshot AI’s publication of Kimi K3.1 Pro weights under a license requiring a separate agreement for Model-as-a-Service businesses above $20 million, and Mistral Large 4’s open weights under a modified MIT license that requires a commercial agreement for Model-as-a-Service businesses. This article compares the four newer releases on two questions a developer meets before adopting a model: what the weights permit, and what the published benchmark numbers represent. Model cards and index pages were read on November 7, 2026; dates in this article are the ones the sources state.

Current situation

Weight access and licenses

Reflection’s Beam model card lists the license as Apache 2.0. The card says Reflection does not prohibit distillation, using outputs to train other models, or fine-tuning, among other listed activities.

The Ripple-65B Preview model card also lists Apache 2.0. VentureBeat writes that the published license contains no revenue or monthly-user threshold requiring a separate commercial agreement, and that Apache 2.0 does not require developers to publish all their modifications under the same license.

Meta’s terms differ. SiliconANGLE reports that the weights come with a new Muse Model License carrying what Meta calls “limited commercial terms”, that no fee applies to developers who adapt the model or ship it inside their own software, and that Meta requires a separate agreement from any company that sells the model as a hosted service and has passed $20 million in gross revenue over a 12-month period. The license text on Meta’s Muse Spark Open model card adds detail. It requires products or services that use the software or any Derivatives to display “Built with Muse Spark Open”, and it defines Derivatives to include any AI model “created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof”. Its Model-as-a-Service clause turns on the aggregate gross revenue of the licensee and its affiliates exceeding $20 million in any consecutive 12-month period, and it does not apply to internal use, research, any commercial activity that is not Model as a Service, or use of the software as a component of the licensee’s own product that is not itself Model as a Service. The attached Usage Policy lists, among other uses that present a risk of death or bodily harm, military and warfare applications and “Operation of critical infrastructure, transportation technologies, or heavy machinery”. SiliconANGLE summarizes this as military and critical infrastructure uses being off-limits.

Cognition’s SWE-2 Flash has no downloadable weights. According to The New Stack, Thinking Machines Lab released Inkling-Small, a 276-billion-parameter model with 12 billion active parameters, on July 30 under Apache 2.0, which lets developers modify and commercialize derivatives without sharing their changes. The outlet writes that keeping SWE-2 Flash’s weights private “sits well within the company’s rights”. Cognition cites cybersecurity: it included cybersecurity tasks in training, raising Flash’s ExploitBench score from 11.3% for Inkling-Small to 34.8%, though The New Stack notes that the benchmark does not establish how the model would perform in real-world attacks or whether releasing the weights would increase misuse risk. Cognition says it is building a program to share weights with select partners and research labs. The outlet reports that it remains unclear who will qualify, what restrictions they will face, or whether they may share models built on SWE-2 Flash. Developers can try the model through Devin Desktop and CLI, where it is free through November 30.

Whose numbers, run how

TechCrunch writes that Reflection’s performance claims for Beam have not been independently verified, and that Reflection’s own benchmarks show Beam outscoring Inkling on four coding tests where both report results, although Inkling is multimodal and Beam is text-only.

The Beam card states that all its evaluations used a temperature of 0.7 and top_p of 0.99, and reports separate scores for BF16, FP8 and NVFP4 numerics. For Terminal Bench v2.1 those are 80.1%, 78.9% and 78.4%; for SWE Bench Pro v2-Hard they are 76.5%, 78.4% and 78.6%. The card marks its technical report as “coming”.

Meta’s card states that Muse Spark Open results used xhigh reasoning effort, a temperature of 1.0, top_p of 1.0 and top_k disabled. In its table the Terminal-Bench 2.1 row is labeled “Terminus 2 agent” (Muse Spark Open 76.4, Beam 80.1), while the DeepSWE 1.1 and SWE-Atlas-QnA rows are labeled “Preferred harness”. SiliconANGLE writes that Chinese open models are stronger at coding, with Kimi K3 leading Terminal-Bench 2.1 at 85 and GLM 5.3 and Qwen3.8 both above Muse Spark Open’s 76.4.

Ricursive’s card says its comparison tables use official model-card scores where available and Ricursive’s own in-house evaluation flow for missing ones. It says it used Terminus 2 for Ripple on SWE-bench Pro v2.0 Hard and DeepSWE 1.1 and OpenCode where a harness was not specified, and that replicating agentic benchmark scores “has been challenging”. The card reports that Terminal-Bench 2.1 scores were universally lower, with the exception of Muse Spark Open, under the official Terminus 2 harness at 128K sequence length with summarization. It lists two Terminal-Bench 2.1 rows:

Model (as named in the card)	“Terminal-Bench 2.1 (self-reported)"	"Terminal-Bench 2.1 (Terminus 2)“
Ripple-65B Preview	80.5	80.5
Reflection Beam 501B	80.1	70.8
Muse Spark Open 816B	76.4	78.7
Inkling 975B	63.8	56.2
Nemotron 3 Ultra 550B	56.4	55.1

By the Herald’s reading of that table, Beam ranks second and Muse Spark Open third in the first column, and the order of those two reverses in the second. VentureBeat writes that Ricursive’s claimed lead applies to selected U.S. competitors and tests, that the company’s own results show it trailing other open models on additional benchmarks, and that the card acknowledges difficulty reproducing agent benchmarks and reports that changing the evaluation framework affected scores.

For SWE-2 Flash, The New Stack reports Cognition’s figures of 20% on Terminal-Bench 4.0 against Inkling-Small’s 0.9%, and a FrontierCode 2.0 score of 24% against 3.2% for the base model. It adds that the company’s own benchmark table shows models built outside the United States, DeepSeek V4.1 Flash among them, outscoring Flash on several tests. Beam, Muse Spark Open and Ripple report Terminal-Bench 2.1 in the cards cited above; the Cognition figures in The New Stack are for Terminal-Bench 4.0.

One independent reference

Artificial Analysis, a third-party evaluator that is not a developer of any of the four models, publishes pages for two of them. Its Beam page gives Beam 1.0 an Intelligence Index score of 18, ranked 19th of 26, against a median of 23 for comparable models, and a cost of $0.28 per Intelligence Index task. Its Muse Spark Open page gives 25, ranked 12th of 26, and a cost of $2.02 per task. Both pages state that the index (version 5.0.1) combines nine evaluations that include Terminal-Bench 4.0 and FrontierCode 2.0; the list also names evaluations such as HLE-Diamond and GDP.pdf. Meta’s card lists an Artificial Analysis Intelligence Index of 39 for Muse Spark Open without naming an index version.

Implications

The points below follow from the texts and reports cited above; they are the Herald’s reading, not forecasts or legal advice.

• By the Herald’s count, the licenses sort the releases into three groups. Two releases list Apache 2.0 (Beam, Ripple-65B Preview). One, Muse Spark Open, carries a label requirement and a Model-as-a-Service clause that, as written, reaches fine-tuned or distilled derivatives and depends on the revenue of the licensee and its affiliates. One, SWE-2 Flash, has private weights, and The New Stack reports that its base license does not oblige Cognition to publish them. The Herald’s earlier reports on Kimi K3.1 Pro and Mistral Large 4 describe Model-as-a-Service conditions in two further licenses.
• A derivative’s weights need not be open because its base’s are. The New Stack describes Inkling-Small as released under Apache 2.0, yet it names Devin Desktop and CLI as the way developers can try the fine-tune built on it. Whether a team can inspect or self-host SWE-2 Flash would depend on a partner program whose eligibility The New Stack says is unclear.
• A benchmark figure carries its setup. The cards cited above show scores that change with numerics (Beam’s three columns), with the harness and sequence length (Ripple’s two rows) and with benchmark version (Terminal-Bench 2.1 versus 4.0). VentureBeat writes that an agent’s surrounding software determines its tool access, long-conversation handling and failure handling, and that a result reflects that setup alongside the model. It also writes that the next step for a development team is a comparison using its own repositories and agent software.
• Independent and vendor measurements answer different questions. Meta’s card puts Beam above Muse Spark Open on Terminal-Bench 2.1 (80.1 versus 76.4), while Artificial Analysis’s index, built from a different set of evaluations, scores Muse Spark Open higher (25 versus 18). The sources cited do not address whether the two orderings conflict; the evaluations behind them differ.

What the sources do not establish

• Among the sources cited here, none reports an independent run of Ripple-65B Preview or SWE-2 Flash on the benchmarks their vendors cite; the Artificial Analysis pages cited cover only Beam and Muse Spark Open.
• Beam’s technical report is marked “coming” on its card, and Meta says a full evaluation report is still to come, according to SiliconANGLE.
• In its November 6 article, The New Stack reports that Cognition says more details of its partner program are coming in “the next few weeks” and that it is unclear who will qualify.
• Model cards and Artificial Analysis pages can change after the November 7 reading; the Artificial Analysis pages carry an “Updated” label without a date.
• Why vendor figures differ is stated only where a card says so; the sources do not explain every gap, including the difference between Meta’s 76.4 and Ricursive’s 78.7 for Muse Spark Open on Terminal-Bench 2.1 with Terminus 2.