Together AI

• No reference anywhere to Indigenous knowledge, data sovereignty, or the CARE Principles across homepage or About page.
• The platform's framing of data is purely operational ('managed storage with zero egress fees'), treating data as a fungible logistics asset rather than as relationship or stewardship.

• No acknowledgment of Indigenous data sovereignty or consent frameworks for training corpora served through inference.
• No consultation with Indigenous communities mentioned for any of the open-weight models hosted.
• No preservation of oral, ceremonial, or non-textual knowledge; the open-model marketplace inherits whatever extractive corpora upstream model-builders used, with Together adding no provenance layer.
• Extractive data practice is structurally invisible: Together is an inference and fine-tuning intermediary, so it disclaims the data-sovereignty question by treating models as already-made artifacts.

Total absence. As an inference/hosting layer, Together AI never engages the question of whose embodied or relational knowledge is encoded in the weights it serves. The intermediary position actively launders the data-sovereignty question rather than addressing it. Floor score.

• Founding date (2022) and named researcher lineage (Tri Dao, Chris Re, Percy Liang, Ce Zhang) give some institutional history.
• Reference to GPU cluster access ('from self-serve instant clusters to thousands of GPUs') gestures at compute political economy without naming it.

• No acknowledgment of colonial data-extraction legacies or the labor history behind training corpora and RLHF annotation.
• No transparency on the geopolitics of GPU access (export controls, NVIDIA dependency) despite NVIDIA being named as a partner.
• No historical humility about AI's inheritances; the narrative is presentist and forward-leaning ('ship faster, scale reliably').
• Regulatory constraints are unmentioned.

A thin institutional self-history exists, and the compute-supply framing brushes against political economy, but there is zero reckoning with the historical processes (extraction, annotation labor, hardware geopolitics) that shaped the stack. Slightly above floor only because dates, lineage and the GPU-supply mention provide minimal anchoring.

• Hosting open-weight models from a global supplier set (MiniMax, Black Forest Labs, Cartesia) implies de facto multilingual and multi-origin model availability.
• Open-source ecosystem framing in principle admits non-Western model contributors to the catalogue.

• No explicit multilingual commitment, language coverage claims, or low-resource-language support beyond incidental catalogue presence.
• No preservation of culturally specific reasoning patterns; models are presented as interchangeable performance commodities ('2x faster inference').
• No consultation with cultural or linguistic scholars.
• Western efficiency logic ('do more with less', 'superior unit economics') is presented as the universal frame of value.

Cross-cultural plurality, where it exists, is an accidental byproduct of being a multi-vendor marketplace, not a stated value. The dominant register is a culturally specific (Silicon Valley) efficiency-and-throughput logic presented as neutral. Low, with a small credit for catalogue diversity.

• Strong, specific, verifiable research output named: FlashAttention series, Mamba-3, BASED, Hyena Hierarchy, speculative decoding.
• Open-weight hosting and open-source contributions enable third-party verification of served models.
• Named academic partnerships (Princeton, CMU, UC Berkeley) provide an external accountability surface.
• Founders are publishing systems researchers, lending the technical claims real provenance.

• Performance superlatives ('2x faster', '60% lower cost', '90% faster pre-training') are stated without linked benchmarks, baselines, or methodology on the page.
• No mention of independent audits of training data or bias for hosted models.
• No third-party replication protocols or known-limitation disclosures for the inference stack.
• Open weights are emphasized for capability, not framed as a verification or accountability instrument.

Genuinely the strongest lens. The systems-research lineage is real, public, and citable, and open weights do permit verification. Capped at 6 because the headline numbers are unsourced marketing superlatives and there is no bias/limitation disclosure or independent-audit posture.

• One affective register exists: 'discover the magic' and 'curiosity-driven discovery' acknowledge wonder and intuition as motivators.
• Hosting generative-media labs (Black Forest Labs image models, Cartesia audio) places creative practice within the platform's ambit.

• No space for ambiguity or poetic uncertainty; the dominant mode is throughput and optimization.
• No recognition of the emotional labor of users, annotators, or creative communities.
• Modes of attention beyond efficiency are absent; 'magic' is deployed instrumentally as developer delight, not as genuine affective inquiry.

A faint affective vocabulary ('magic', 'curiosity') and a generative-media catalogue earn it above the floor, but these are conscripted into a performance narrative rather than opening real space for feeling, ambiguity, or the felt texture of the work.

• 'Model Stewardship' and 'building everything with the purpose of benefiting society' assert a future-facing governance intent.
• Efficiency framing ('do more with less') implies, but never states, an environmental-cost awareness.

• No engagement with labor-displacement risk from the agentic and inference systems being scaled.
• No environmental or energy-cost disclosure despite operating GPU fleets 'to thousands of GPUs'.
• No democratic or participatory governance of the agentic systems Together helps deploy; governance is asserted as a corporate value, not a shared process.
• 'Benefiting society' is unoperationalized, with no mechanism, metric, or accountability attached.

The stewardship and 'benefit society' language is the lab's clearest responsibility gesture, but it is entirely unoperationalized, and the two material future-shaping questions for an inference provider, energy footprint and labor displacement, are simply absent. Low.

• 'Democratized AI access' and 'self-serve instant clusters' lower the cost-of-entry barrier, which can incidentally widen access for smaller and Global-South developers.
• Open-source community framing nominally invites broad contribution.

• No participatory design with Global South developers; access is transactional (pay-per-token), not participatory.
• No disability-community accessibility commitment.
• No labor-representative engagement (data annotators, content moderators in the upstream supply chain).
• No compensated feedback channels; 'community' means open-source users, not represented stakeholders.

Cost-democratization is real and not nothing, but it conflates cheaper market access with inclusion. None of the named marginalised constituencies, disabled users, annotation labor, Global-South co-designers, has a seat or a compensated channel. Low.

• The catalogue model creates one self-aware-adjacent tension Together could name but does not: it brands itself on 'openness' and 'responsibility' while functioning as a neutral hosting rail that disclaims responsibility for what the weights it serves actually contain.
• 'Together' as a name against a product whose core promise is 'eliminating infrastructure management', i.e. removing humans from the loop, is an unremarked irony.

• No willingness to name the central contradiction of an 'open and responsible' intermediary that adds no provenance, safety, or data-rights layer to the models it monetizes.
• No irony, satire, or paradox deployed as instrument; the register is uniformly earnest and promotional.
• The lab's own seriousness ('discover the magic', 'benefiting society') is treated as exempt from audit.
• No space where the official narrative is tested by its opposite.

Polished, frictionless, and entirely un-self-interrogating. The richest available inversion, an 'open/responsible' brand that is structurally a responsibility-disclaiming pass-through, is left untouched. Near floor; a point above only because the contradiction is so legible it half-surfaces itself.