Baseten

Thesis

Baseten is deploying its $1.5B Series F across three fronts at once: a capital-intensive compute-supply function, an enterprise/commercialization layer, and an in-house post-training research arm (“Base Labs”) that now ships open models. The hiring, release, and fork signals point to a neocloud moving beyond raw inference capacity toward differentiated open models, RL-environment services, and enterprise governance — not just GPU resale. P4P26P9P12P11W4

Signal desks

Hiring

• Compute: “Strategic Compute Investments” (SF) — structures, models, and negotiates multi-year capacity, data-center, and powered-land transactions with the Head of Supply, Finance, and leadership; framed as the “most capital-intensive and strategically important” function. P4E3
• Product Platform is the densest hiring cluster: founding Identity & Authorization role (fine-grained authz for users, service accounts, and agentic workloads) P11E14; Billing & Internal Tooling (Orb integration) P12E13; Partner Platform (model labs, hyperscalers, dev platforms) P13E12; Backend Engineer (telemetry/observability pipelines, billing, internal AI agent platform) P14E11; Product Engineer (enterprise controls, agentic development) P20E17.
• Applied research: Post-Training Applied Researcher (EPD, SF, $200K–$275K + equity). W1
• GTM/scale: Technical Program Manager, Forward Deployed Engineering P19E16; Product Marketing Manager, Partnerships E33; Partner Enablement Lead E39; Sales Manager – Enterprise (New York) E47.
• Infra/ops: Engineering Manager, Runtime – Storage E40; IT Systems Engineer (listed Remote; description notes hybrid SF/NY) P18E20; Finance: Global Equity Programs Lead (equity administration across countries/entities) P5E8.

Forks

• action-junit-report — fork of mikepenz/action-junit-report (Apache-2.0; 6 open issues); CI test-reporting tooling. P2
• run-report-action — fork of moonrepo/run-report-action (TypeScript); CI run-report tooling. P3P1
• sandbox — fork of blaxel-ai/sandbox; sandboxing runtime. E52
• harbor — fork of goharbor/harbor; container registry. E60
• Interpretation: forks cluster on CI/DX plumbing plus container/sandbox infrastructure, consistent with the Carbon/OpenShell sandbox work. P2P3E52E60W5

Releases

• SDK/CLI wave: baseten-python v0.13.1 (sandbox filesystem memory) P6E7; baseten-go v0.5.0 (sandbox previews) P28E30; baseten-js v0.4.0 E28; baseten-cli v1.5.0 (new “baseten grid” agent-harness routing command group) P15E15; truss v0.18.36 (batch System One API, ml-cookbook-resolved train examples) P24E18.
• Security-forced rebuilds: baseten-cli v1.4.1 P22E23, truss v0.18.36 P24E18, baseten-bdn v0.1.2 P21E21, and baseten-switch v0.7.2 P27E24 were all rebuilt/re-pinned after the eMudhra/emSign CA key compromise. P16E9
• Model/product: elephant-1 and mammoth-1 decision models live via the System One endpoint P25E22; baseten-switch v0.8.0 (beta) routes Claude Desktop to Baseten models. P7E2

Talking

• Decision models: “the top ranked decision models in the world” — Base Labs-trained mammoth-1/elephant-1. P26E1
• Post-training services: Baseten Environment Network of vetted RL-environment partners. P9E5
• Security/compliance: automatic API-key revocation (Nov 5) P8E6; 401-vs-403 auth error change P10E4; CLI security notice P16E9.
• Open-weight safety stance: Base Labs partnership with Hugging Face and Goodfire for safety eval/monitoring of open-weight models. W4
• Infra/sandbox narrative: Carbon with NVIDIA OpenShell and Blaxel sandboxes. W5
• Observability: GPU temperature, tensor-core utilization, and memory-bandwidth metrics. P17E10

Shipping

• Decision models shipped to the Model API: elephant-1 (9B) and mammoth-1 (27B) via the System One endpoint, with a working curl example and model IDs decision-models/baseten-elephant-9b-latest / decision-models/baseten-mammoth-27b-latest. P25E22
• baseten-cli v1.5.0 ships “baseten grid” (pre-release) for agent-harness routing — setup/sync/status/teardown, route management, team/user spend limits, provider connections, and managed classifiers. P15E15
• baseten-python v0.13.1 adds sandbox filesystem memory control (max_memory_usage). P6E7
• baseten-go v0.5.0 adds sandbox previews (HTTPS, public or token-gated) and image push memory/volume sizing. P28E30
• truss v0.18.36 adds a batch System One API and resolves train init --examples through ml-cookbook; Baseten ships release candidates (v0.18.36rc2) before stable. P24E18P23E19
• Security rebuilds shipped in the same window: baseten-cli v1.4.1, truss v0.18.36, baseten-bdn v0.1.2, baseten-switch v0.7.2. P22P24P21P27P16
• baseten-switch v0.8.0 (beta): Claude Desktop gateway mode, model-picker management, and config backups; Homebrew is the supported install channel. P7E2

Research themes

• Decision models as a new category: post-trained Qwen3.5-9B and Qwen3.5-27B into classification/routing/scoring models; mammoth-1 ranks #1 of 208 on the Jev Decision Index. P26
• RL environments and post-training: in-house “science around RL environment creation” now scaled through a vetted partner network (engagements with Harvey, Lovable, Gamma). P9
• Open-weight safety and interpretability: Base Labs partnership with Hugging Face and Goodfire targets safety eval/monitoring and the abliteration problem; Baseten Labs traces to the Parsed (mechanistic interpretability) acquisition. W4W6
• Agent/coding efficiency: DeepSeek-V4.1-Flash (552B multimodal MoE, Causal Encoder-Decoder, improved KV cache/prefill) W2; Reflection Beam (501B MoE, 23B active, 1M context, Apache 2.0). W3
• Sandboxing/security infra: Carbon ships kernel capabilities for agents, manual snapshotting, and forking. W5
• Replication and cookbooks: talkie-facts-replication (sequential fact-writing experiments) E54; ml-cookbook (ready-to-use ML training recipes). E25

Hiring & scaling

• Hiring is heavily San Francisco-concentrated; the exceptions are a New York enterprise sales role and a Remote IT role. P4E47P18E20
• Two distinct scale narratives coexist: enterprise/GTM buildout (billing, identity, partnerships, sales, enablement) and research/supply buildout (post-training researchers, compute investments). P12P11P13E33E39E47W1P4
• Compute supply is framed as a capital-markets function: underwriting, TCO/return modeling, and negotiation of large compute/DC transactions. P4
• The Global Equity Programs Lead signals multi-entity/international scaling and growing equity-admin complexity. P5
• Forward Deployed Engineering is described as “scaling the team quickly,” now adding a TPM “operating system.” P19

Category implications

• Strategy: Baseten is positioning itself beyond inference-as-a-service into differentiated open models and post-training services (decision models, RL-environment network), cooperating with the open-weight frontier rather than only serving it. P26P9W4
• Infrastructure: capital-intensive compute sourcing plus multi-cloud observability/telemetry pipelines imply sustained GPU/DC capex and platform-observability spend. P4P14
• Product: enterprise governance is becoming a first-class product surface — fine-grained authorization, billing/metering (Orb), partner integrations, and agent-harness routing (“grid”). P11P12P13P15
• Research: in-house Base Labs plus the partner network is a research-as-a-service moat tied to open-source models and open-weight safety. P26P9W4
• Hiring: the role mix maps to enterprise GTM and supply-chain finance, not just core inference engineering — a maturity signal. P12P11P4E33E39
• GTM: partner enablement/marketing roles, an enterprise New York sales seat, and a changelog banner offering “OpenAI commitments on Baseten open models” suggest a partner/marketplace-led motion. E33E39E47P25

Traction highlights

• $1.5B Series F led by Altimeter Capital, Conviction Partners, and Spark Capital. P4
• Named customers: Cursor, Notion, OpenEvidence, Abridge, Clay, Gamma, and Writer. P4
• mammoth-1 is #1 of 208 decision models on the Jev Decision Index (0.4 score 71.24); elephant-1 is the top 9B model; Artificial Analysis scores are 65.9 (mammoth-1) and 65.2 (elephant-1). P26
• The mammoth-1/elephant-1 blog drew HN attention (8 points). E1
• Post-training engagement customers include Harvey, Lovable, and Gamma. P9
• Enterprise authorization customers cited include Harvey, HubSpot, and Notion. P11
• Popular-model catalog lists mammoth-1, elephant-1, Muse Spark Open, DeepSeek V4.1 Flash 1017, and GLM-5.3 variants. P8