Muse Spark Open vs SWE-2 Flash

Muse Spark Open 69

Model reading of the checklist (tool use, structured output, caching, context, throughput, defaults and SDKs, errors). OpenAI-format tools and parallel calls come back as tool_calls, and the template supports namespaced tools and deferred loading through a tool_search tool. A forced tool_choice is not documented (17 of 20). JSON mode and JSON Schema come from each runtime's guided decoding, which Meta's docs show for TensorRT-LLM only (8 of 15). Prefix caching is the runtime's. Meta's SGLang command turns on hierarchical caching, and Dynamo routes each request to the worker holding its cached context (8 of 10). 1,048,576-token context, with sliding-window attention on 46 of 62 layers holding KV memory down (15 of 15). Draft models for speculative decoding in SGLang and vLLM, and no batch mode of its own (7 of 10). Any OpenAI SDK works against a local server, but reasoning can't be turned off, defaults to the most expensive level, and needs streaming, a max_tokens budget for reasoning and the right stop tokens (6 of 15). The template raises named errors for an unsupported reasoning effort, tool arguments passed as a string and a tool message with no name. Server errors are the runtime's (8 of 15). 17 + 8 + 8 + 15 + 7 + 6 + 8 = 69.

SWE-2 Flash 10

Model reading (tool use, structured output, caching, context, batch, SDKs). Tool use is Devin's harness, not something a caller configures; no context length, output limit or cache price is published for SWE-2 Flash (5 of 25). No documented output-size or reasoning controls for this model; SWE-2's effort levels aren't stated for Flash (0 of 20). No model-level error documentation (0 of 20). No idempotency or retry guidance (0 of 20). One flag or picker selects it and devin -p runs a single turn and prints to stdout, but there is no SDK or direct endpoint (5 of 15). 5 + 0 + 0 + 0 + 5 = 10.

Muse Spark Open 61

Read for a model the owner serves. No specification of its own. The contract is the Jinja chat template, which renders OpenAI-format tools with additionalProperties: false added and rejects bad values, plus the OpenAI-compatible servers in vLLM and SGLang with the muse parsers (10 of 25). dev.meta.ai/llms.txt lists the Muse Spark Open overview and seven pages under it, plus three cookbook pages, each with a Markdown copy (10). The card lists seven intended uses and six limitations, the docs say when to pick BF16 or NVFP4 and which reasoning strength suits which task, and the Usage Policy states what is out of scope (17 of 20). reasoning_effort is an enum of four values and anything else returns an error, and tools take JSON Schema. Structured output is documented only for TensorRT-LLM's response_format, not in the prompting guide or for vLLM and SGLang (11 of 15). curl and Python examples for every runtime and hosted provider, the rendered prompt for each case, and the template's own error messages, but no error reference for a served model (11 of 15). One release, with no changelog or version scheme beyond the licence's version line, Muse Spark Open (2 of 15). 10 + 10 + 17 + 11 + 11 + 2 = 61.

SWE-2 Flash 29

Model reading. No model API, so no OpenAPI entry; the Devin v3 OpenAPI file lists SWE-2 agents only, and no model card was found (0). docs.devin.ai serves llms.txt and Markdown copies of every page (10). The launch post says what it is for (automations, code review, feature development, documentation) and the CLI docs say when to pick a SWE model, but the models pages don't list SWE-2 Flash yet (6 of 20). No documented model ID, reasoning settings, context or output limits to type (0 of 15). The CLI pages show --model and /model examples, with no model-specific errors (3 of 15). Named, dated model versions on Cognition's blog and public release notes, but no release note for SWE-2 Flash yet; the newest is 2 November (10 of 15). 0 + 10 + 6 + 0 + 3 + 10 = 29.

Muse Spark Open 50

Self-hosted rule. Nothing to buy from Meta, and the licence grants use free of charge. The Model as a Service clause asks for a separate agreement only from licensee groups above $20M gross revenue in 12 months (20). Free to run with no card (20). Meta's own weights need a manually reviewed Hugging Face form with legal name and date of birth, or, since the 4 November re-read of Get the model, a signed-in download from Meta's own page, so an agent can't get them alone. The ungated GGUF quantisations from ggml-org and Unsloth that Meta's docs point to download without approval (the Unsloth repository is ungated on 4 November), so part of the line (10 of 20). No payment protocol (0). Hosted copies at $1.25 input and $4.25 output per million tokens need each provider's own account and are not graded. 20 + 20 + 10 + 0 = 50.

SWE-2 Flash 20

Published rubric. No payment protocol (0). Plan prices are public, but no per-token price for SWE-2 Flash is on a readable page; the docs' cost table loads in the browser and the post gives only 'one-fifth the price' of SWE-2 (10). Free through 30 November 2026 on Devin Desktop and CLI, but the docs say Free tier users only have SWE-1.6, so whether the Free plan reaches it, and without a card, isn't confirmed; the in-between value used on the devin listing (10). A person signs in through a browser (0). 0 + 10 + 10 + 0 = 20.

Muse Spark Open 53

Weights are downloadable under the Muse Model Licence, which is clear and broad in its grant but not an OSI licence. It adds a display requirement, a separate agreement before Model as a Service by a licensee group above $20M gross revenue, and a Usage Policy with field-of-use limits (military, nuclear, espionage, critical infrastructure, heavy machinery) and an age limit of 18. It has no regional exclusion and no governing-law clause. Training code and data aren't released (18 of 30). Inputs stay on the owner's hardware, and Meta's docs say self-hosting keeps 'every prompt and response' inside the owner's infrastructure. The access form's details, and a download from Meta's own page after signing in, go to Meta under its general privacy policy, which we didn't read, and the training data is described in outline only (20 of 30). No deprecation policy, dated notices or versioning found (0 of 20). The Get the model page says everything runs offline once the weights land, and safetensors weights carry no code. No other telemetry statement, and Meta's download form records name, date of birth, country and IP location (12 of 20). 18 + 20 + 0 + 12 = 50.

SWE-2 Flash 64

Editorial half. Closed model under Cognition's platform terms; the weights aren't public and the sharing programme isn't described yet (15). Terms, privacy policy, security page and DPA agree, with training on self-serve data by default and no retention periods, as read for the devin listing (18). The API versioning policy covers API endpoints, not models, and no model deprecation notice was found (5 of 20). Sub-processors listed with purpose and country, per the devin listing's 8 October check (18). 15 + 18 + 5 + 18 = 56.

Muse Spark Open 44

Read as software the owner runs. No account or key to run it. Meta's weights need manual approval on Hugging Face and an access token, and server authentication is the owner's to add (12 of 30). An agent model that calls whatever tools the operator gives it. The card recommends human-in-the-loop confirmation for irreversible actions and Meta reports agentic-risk evaluations, but nothing in the model or template enforces an approval (8 of 20). The chat template neutralises special-token sequences and media placeholders in caller text and wraps tool results in <tool_output>, and the card advises protecting internal assistant messages and layering defences against adversarial input. No indirect prompt-injection results are published (10 of 15). No call log of its own. Visibility depends on the serving stack (5 of 15). Weights are safetensors that load through Transformers' own classes from 5.19, and Meta published cyber, chemical, biological and misalignment evaluations. dev.meta.ai/.well-known/security.txt answers 404 (re-read on 4 November), research.meta.ai per the 15 October check, and no security route is on the card or in the docs, only museoss@meta.com and the Hugging Face discussions (9 of 20). 12 + 8 + 10 + 5 + 9 = 44.

SWE-2 Flash 53

Read as a hosted model reached through Devin's local clients. The CLI signs in through a browser, or the enterprise identity provider, and stores an API token in credentials.toml that by default is persistent and does not expire and carries the user's access, with no scopes found for it (10 of 30). Devin CLI documents an OS-level sandbox with network filtering, hooks that can block tool calls, and enterprise model allowlists and team settings (14 of 20). Guardrails against prompt injection are Enterprise only and cover user messages, as on the devin listing (8 of 15). Enterprise analytics count tool calls and consumption by model and user; no per-user audit log for the CLI was found (8 of 15). SOC 2 Type II and ISO/IEC 27001:2022, a disclosure address, no bug bounty and no security.txt, per the devin listing (13 of 20). Training: self-serve customer data may train Cognition's models by default, with an opt-out on paid plans that also turns on zero data retention at model providers. Cognition withheld the weights over cyber misuse, which is recorded as a fact, not scored. 10 + 14 + 8 + 8 + 13 = 53.

Muse Spark Open 42

Read with the local-software lines, since the owner serves the weights. Meta's weights come from two Hugging Face repositories gated with manual review. The docs state Python 3.10 or later, Transformers 5.19 or later, Docker images for SGLang and vLLM and the GPUs for each checkpoint (4 × B300 for NVFP4, 8 × B300 for BF16), but SGLang's image is a dev tag and TensorRT-LLM's support sits on a feature branch (13 of 20). No CI or test suite of Meta's for the model was found. github.com/meta-models answered 403 to our fetcher, so its repositories are unchecked (0 of 25). On 4 November the BF16 repository's discussions page lists 13 open threads from 15 October to 1 November, 10 discussions and 3 pull requests (one a draft). A script counted 29 replies across the seven threads with replies, six discussions and the draft pull request, none from the repository owner or a Meta member, and their text wasn't read, so crash reports weren't separated from other posts (15 of 25, held from 15 October and flagged). No changelog, version tags or revision notes, and the commit history needs gated access (2 of 15). Released as a final checkpoint with no preview label from Meta, and both repositories unchanged since 15 October, though runtime support ships as launch-specific images and a branch, and Groq lists it as a preview (12 of 15). 13 + 0 + 15 + 2 + 12 = 42.

SWE-2 Flash 50

Model reading, for the surface it runs on, Devin Desktop and Devin CLI. Status page at www.devinstatus.com (Atlassian Statuspage) with Desktop Agent and Desktop Tab components besides the cloud ones, read on 4 November 2026 (20). The history feed for the 90 days from 6 August lists 'Devin unavailable' for 20 minutes on 13 August with no component named, read as across the board and a major as on the devin listing; Devin Cloud unavailable on 13 August and the web app down on 24 September name cloud surfaces this model doesn't use; the rest are degraded-performance incidents (new sessions on 3 September, 'multiple services' on 17 September, Ask Devin on 26 October, automation triggers on 4 November) and session queuing on 10 and 11 August, read as minor and a scheduled Devin Desktop extension marketplace upgrade on 7 August, read as maintenance. One major (10). No rate limit figures for the model or the CLI; usage is a daily and weekly quota with no numbers (0). No 429, backoff or retry guidance for this surface (0). Cognition publishes an SLA (25 February 2026) with uptime and response times that applies when included in an order form, read as a contractual SLA on Enterprise, as on other listings (10). The launch post makes it available today with no preview or beta label (10). 20 + 10 + 0 + 0 + 10 + 10 = 50.