Provenance Record

Clear, neutral News article of about 770 words (script count; within the 400-1200 News range). Structure separates the license terms, checkpoints, pricing, vendor benchmark claims, independent measurements, a conflicting-figures section and unknowns. Title is 137 characters (under 150).

Read all six gzipped snapshots (source-0 to source-5) as raw extracted text; none is a failed snapshot and none is an Archive.org fallback. The VentureBeat snapshot (source-5, status 200) is full raw text, so the writer's WebFetch-summary caveat is resolved here by direct reading. source-0 mistral.ai post: confirms 'Mistral Ecosystem Open-Weight license, a modified MIT license', 'free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially'; pricing $0.70 input / $0.07 cached input / $2.10 output per M tokens (also in the pricing widget); FrontierCode 2.0 19.7%; Terminal-Bench 4.0 up 13 points to 41%; Coding Agent Index 48% effectively tied with GLM-5.3 with 47% fewer tokens; Cyber Index 51%; 96.6% average refusal 'the highest we measure among open models'; availability in Studio, API, OpenRouter, NVIDIA-hosted endpoints and NIM container, Microsoft Foundry, weights on Hugging Face; 'Earlier this month' preview; guardrails being modified. All confirmed. source-1 model card: confirms ~1T total, 128 routed experts with 4 active, 1M context, NVFP4 checkpoint, DSpark draft head with 'up to 1.9x higher throughput than serving the target model alone. Actual gains depend on hardware and workload', 'leading open-weight models from China score higher on this aggregate index'. The card lists FP8 only as a repository tag (the blog states two official variants, FP8 and NVFP4), a trivial attribution looseness. The card states 49B active per token (52B including embeddings and output layers) in its introduction and ~52B in Key Features, so the article's active-parameter 'conflict' is not a conflict (correction filed). source-2 LICENSE: read in full. Permission grant covers use, copy, modify, merge, publish, distribute, sublicense and sell, run, deploy, further train; Section 2 requires a separate commercial license before using the Model or derivatives for any commercial purpose if the Licensee or any affiliate operates a Model as a Service business; the definition matches the article, and it also carves out end-user products with embedded capabilities and mere relaying of requests; Section 3 exempts internal use and use accessed through Mistral's official products. The title's 'modified MIT license' and 'requires a commercial agreement for Model-as-a-Service businesses' are supported by the license text and Mistral's post, and the article does not characterise the license beyond its text (the article omits the Section 2 carve-outs, which is a completeness point, not an error). source-3 Hugging Face blog (published October 29, 2026): 'about 1T total, 49B active', 1M context, two official variants FP8 and NVFP4, GGUFs by Unsloth and the llama.cpp team, 'MIT derivative that establishes limitations for model-as-a-service offerings'. Confirmed. source-4 Artificial Analysis: Intelligence Index 26, median 23 for comparable open-weight models, $0.70 in / $2.10 out, $2.06 per Intelligence Index task, total parameters 1.1T, active 52B, 1M context, 'was released on October 29, 2026'; also labels the model 'Open Weights (Commercial Use Restricted)'. Confirmed. source-5 VentureBeat (Carl Franzen, October 29, 2026): 26 vs 25 for the preview, six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3 (32), $2.06 vs $5.54 per task (63%), hallucination 28% vs 41%, 'custom license', and the Janiewicz quote 'I don't want to get disconnected. I want to own my model,' appears verbatim; she is identified as Mistral's chief revenue officer and the interview was in person earlier that week. All confirmed. VentureBeat also reports the earlier preview API rates as $1.36 / $0.14 / $4.18, matching the Herald preview article. Herald preview article opened (src/content/articles/2026-10/07-mistral-releases-large-4-in-public-preview-...): it reports 49 billion active parameters and API pricing of $1.36 input / $4.18 output from the Mistral announcement, exactly as described, and its headline says open weights were 'promised'; the internal link resolves to an existing article. Prompt-injection scanner: the manifest flagged four sources, all judged false positives after reading the matches in context. source-0 (system-prompt-reference): the matched text is 'agents shouldn't open you up to threats like system prompt exfiltration or malicious code injection', Mistral's description of its agent-security benchmark, not an instruction to a reader. source-1 and source-2 (system-prompt-reference and chat-template-marker): the match is the model's embedded chat_template_jinja metadata, including '{#- Default system message if no system prompt is passed. #}' and the literal '{{- '[INST]' -}}' template tokens, which is template code in the repository page, not directed at an AI reader (the LICENSE page snapshot carries the same repository metadata). source-3 (system-prompt-reference): 'The fine tuned model would be able to reason in the language the system prompt asks for!' is a fine-tuning tutorial sentence. I searched all six snapshots for 'ignore previous/prior/above', 'do not tell the user' and text addressing an AI agent and found no instruction aimed at a reader; the other 'AI agent' hits are site navigation and article titles. Scanner auto-REJECT overridden on that basis. Independence: Mistral (source-0) is the subject; Hugging Face (sources 1-3) hosts the weights and is named by Mistral as a distribution partner (its VP Jeff Boudier is quoted in Mistral's post), so it is not independent of the subject; Artificial Analysis (source-4) is an independent benchmark provider with no stake in Mistral, although Mistral cites its indices and some AA pages cover the same release, so it counts as independent with that caveat; VentureBeat (source-5) is an independent outlet that interviewed a Mistral executive. Independent-source count: 2 of 6 (Artificial Analysis, VentureBeat). The article is News, so the Analysis minimums do not apply.

Headline, summary and lead are fully supported (license terms, Artificial Analysis score of 26). Pricing is correctly dated 'as read on November 4, 2026' (snapshots fetched that day) and correctly contrasted with the preview's $1.36/$4.18. Mistral's benchmark figures are consistently labelled as Mistral's own. Two subordinate errors are filed as corrections: (1) the 'What We Don't Know' bullet saying no independent replication of the FrontierCode, Terminal-Bench or cyber figures was found is contradicted by the cited VentureBeat article, which reports Artificial Analysis's own results (Terminal-Bench 4.0 25% vs Mistral's 41% in different evaluation contexts; FrontierCode 2.0 15% vs Mistral's 19.7%, with DeepSeek at 29% and GLM-5.3 at 22%; Cyber Index 51 and Coding Agent Index 48 confirmed) and says the independent evaluation does not support extending Mistral's FrontierCode first-place claim; (2) the 'Conflicting Figures' framing of 49B vs 52B active parameters misses that the model card itself reconciles them (49B per token, 52B including embeddings and output layers), as VentureBeat also reports. The article's Mistral-attributed 'Cyber Index 51%' is in fact confirmed by Artificial Analysis. The article's phrase that VentureBeat describes the license as custom 'rather than a standard open-source one' is the writer's gloss on VentureBeat's 'custom license' and 'modified MIT license with an additional commercial restriction'; acceptable. The writer-reported 975-word count differs from the script's 767; both are within range.

Substantively accurate, well-sourced license coverage with two recoverable errors in subordinate sections, published with a public corrections record. Featured candidacy: 2 - an open-weight release from a major lab with a notable commercial-use condition, relevant to a broad AI audience.