What Goodfire does

Goodfire is a San Francisco-based AI interpretability research lab and U.S. public benefit corporation, founded in June 2024, that builds "Silico," a mechanistic interpretability platform for inspecting, debugging, and intentionally designing frontier foundation models. The platform combines sparse autoencoders, mechanistic interpretability, neural geometry analysis, and AI-agent-driven 'model neuroscientist' workflows, and is delivered through three vertical SKUs targeting life sciences (Mayo Clinic, Prima Mente, Arc Institute), robotics and computer vision, and large language models (Gemma 3, DeepSeek-R1, GPT-OSS, Llama). Underlying capabilities include named proprietary techniques such as RLFR (Reinforcement Learning from Feature Rewards), covariance-based sequence pooling for variant-effect prediction, probe-guided self-correcting search via MALA, and attention-probe early-exit for chain-of-thought — results from which have appeared in Nature and on arXiv and have produced the publicly available EVEE resource covering 4.2 million ClinVar variants.

The company monetizes through gated enterprise platform licensing (Silico access via 'Request early access'), verticalized research engagements with named partners, and usage-based compute on its frontier-scale interpretability infrastructure that harvests activations from trillion-parameter models. Customer access is sales-led and quota-based, with named enterprise customers including Microsoft, Rakuten (production PII detection since October 2025), Mayo Clinic, Arc Institute, and Prima Mente. Pricing is not publicly disclosed; access is gated behind a contact form. The company has raised over $200 million across three rounds — a $7M seed (August 2024, Lightspeed), a $50M Series A (April 2025, Menlo Ventures, including a $1M investment from Anthropic — its first startup investment), and a $150M Series B (February 2026, B Capital) at a $1.25B post-money valuation. Goodfire operates in person five days a week in San Francisco with a secondary New York location, employs 11-50 people, and holds SOC 2 Type II certification as of May 2026.

Name

Goodfire

Legal name

Goodfire

Website

https://goodfire.ai

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

11–50 employees

Short description

Goodfire is a San Francisco-based AI interpretability research lab that builds Silico, a mechanistic interpretability platform helping enterprises and research institutions understand, debug, and intentionally design frontier foundation models across life sciences, robotics/vision, and LLMs.

Ownership category

akta.pro rank

Product category

AI Interpretability Software

NAICS

Computer Systems Design Services (541512), Other Computer Related Services (541519)

SIC

Services-Prepackaged Software (7372), Services-Computer Programming Services (7371)

akta.pro primary industry

Foundation Model Developers (LLM/Multimodal Model Labs) (HDAAACAA)

akta.pro secondary industries

Evaluation, Benchmarking & Observability (Evals, Monitoring, Tracing) (HDAAACAH), Bioinformatics & Multi-omics Analysis Software (HLAGAJAD), Safety, Alignment & Content Moderation (Guardrails, Red Teaming) (HDAAACAI), Safety & Alignment Evaluation (red-teaming, harmful capability testing) (HDAAAMAL)

Keywords

• AI interpretability platform
• Mechanistic interpretability software
• AI debugging tools
• Neural network analysis
• Foundation model evaluation

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices2 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Marketing or Sales, Infrastructure

Revenue model

1. Enterprise Platform Licensing (Silico): Revenue generated through enterprise contracts with Fortune 500 companies, major healthcare institutions, and AI research labs that 'request early access' to the Silico platform. Model is gated behind a sales conversation rather than self-serve, indicating recurring platform/SaaS contracts likely sized per use case and compute consumption.
2. Verticalized Research Engagements: Custom research partnerships and applied interpretability work with named partners (Mayo Clinic for genomics, Prima Mente for epigenetics, Arc Institute for Evo 2, Radical AI for materials, Rakuten for production PII detection), often generating jointly authored research outputs that also serve as reference customers.
3. Compute/Usage-Based Inference on Silico: Operating model includes hosting frontier-scale activations (e.g., 'harvesting activations from a trillion-parameter model') and serving an interactive EVEE web resource for all 4.2M ClinVar variants, implying usage/compute-based monetization alongside base platform subscriptions.

Pricing tiers

Go-to-market motion3 records

Distribution channels4 records

Marketing channels9 records

Core offering

Goodfire builds Silico, a mechanistic interpretability platform that decomposes foundation models into interpretable features so teams can understand internal representations, run health checks, debug failures, and intentionally shape AI model behavior. The platform is delivered in three vertical SKUs — Silico for Life Sciences, Silico for Robotics & Vision, and Silico for LLMs — and combines sparse autoencoders, mechanistic interpretability, neural geometry analysis, and AI-agent-driven 'model neuroscientist' workflows.

Product overview

Goodfire offers a single unified interpretability platform — Silico — organized as a core platform plus three vertical sub-products tailored to model modality. The flagship Silico platform ("the first platform for intentional model design") combines frontier interpretability techniques, an AI "model neuroscientist" agent, and a model design environment on frontier-scale infrastructure; teams use it to see inside every prediction, run health checks, debug failures, and shape model behavior. Silico is delivered in three verticals: Silico for Life Sciences (used to interpret Mayo Clinic and Prima Mente genomics models and produce the public EVEE resource), Silico for Robotics & Vision (validated on robotics policies and the EchoJEPA cardiac vision model), and Silico for LLMs (validated on Gemma 3 12B, DeepSeek-R1, GPT-OSS-120B, Llama 3.1/3.3 and powering RLFR hallucination reduction). The Ember platform, introduced with the April 2025 Series A, is the predecessor name now subsumed under Silico. Access is gated via partnership, with early-access requests submitted through the company's contact page.

Differentiator

Problem solved

Functional benefit

Brands

• Silico: The company's flagship platform for intentional model design, allowing developers to build, understand, debug, and intentionally design AI models with the precision of written software.

Products and services

• Silico Goodfire's flagship mechanistic interpretability platform — 'the first platform for intentional model design.' Combines frontier interpretability techniques, an AI 'model neuroscientist' agent, and a model design environment on frontier-scale infrastructure so teams can see inside every prediction, run health checks, debug failures, and shape model behavior.
• Silico for Life Sciences Life-sciences vertical of Silico. Validates what biological models actually learned, unlocks new science (biomarkers, mechanistic hypotheses) from model internals, and designs custom biological models built on real understanding.
• Silico for Robotics & Vision Robotics and vision vertical of Silico. Catches generalization failure before deployment, traces failures to specific training sequences, and fixes physical behavior without retraining by steering between latent policy modes.
• Silico for LLMs Large-language-model vertical of Silico. Anticipates failures before deployment, corrects failure modes directly without retraining, and precisely controls what the model is learning through probes, RLFR, and probe-based data attribution.
• Evo Variant Effect Explorer (EVEE) Interactive web resource delivering pre-computed pathogenicity scores, disruption profiles, and on-demand natural-language explanations (synthesized by Claude Sonnet 4.6) for all 4.2 million variants in ClinVar, built on top of Evo 2 embeddings and Goodfire's annotation probes.

Quantifiable outcome

• 58% reduction in hallucinations on Gemma-3-12B-IT using RLFR, at ~90x lower cost per intervention than LLM-as-judge with no degradation on standard benchmarks
• +6 more outcomes

Named customers5 records

Segments5 records

Ideal customer profiles5 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability15 records

Feature7 records

Partnerships

Five partnerships are on record, tiered flagship and strategic.

• Prima MenteflagshipStrategic or Co-development PartnerMulti-paper research collaboration that produced the January 28, 2026 paper 'Using Interpretability to Identify a Novel Class of Biomarkers for Alzheimer's Detection' (Wang et al.). Goodfire applied Silico to interpret Prima Mente's Pleiades epigenetic foundation model, discovering that DNA fragment-length patterns dominate AD detection — the first major natural-science finding obtained from reverse-engineering a foundation model. Joint Goodfire/Prima Mente team of researchers listed as authors. Prima Mente is also a named customer using Silico.
• Mayo ClinicflagshipStrategic or Co-development PartnerStrategic research collaboration to interpret the Evo 2 genomic foundation model for variant effect prediction. Joint research produced EVEE, an interactive web resource covering all 4.2M ClinVar variants, achieving state-of-the-art performance (0.997 AUROC on 839k ClinVar variants, 0.991 AUROC zero-shot on indels). Joint Goodfire/Mayo Clinic team including Joshua Meehl, Carl Molnar, Eric Klee, and others. Mayo Clinic is also a named customer on the Goodfire homepage.
• Arc InstituteflagshipStrategic or Co-development PartnerResearch collaboration interpreting Arc Institute's Evo 2 genomic foundation model, with results published in Nature. Decomposed Evo 2 into interpretable features spanning sequence, structure, and function, including finding the tree of life as a manifold. Arc Institute is a named customer/partner on the Goodfire homepage.
• Radical AIstrategicStrategic or Co-development PartnerPartnership announced July 30, 2025 ('Partnering with Radical AI to Advance Materials Science with Interpretability') and producing the April 1, 2026 paper 'Using Self-Correcting Search to Accelerate Materials Discovery.' Goodfire applied probe-guided self-correcting search to Radical AI's MatterGen diffusion model, yielding ~30% more viable candidate materials in a target band gap range.
• Genomic Medicine Collaboration PartnerstrategicStrategic or Co-development PartnerCollaboration announced September 9, 2025 to advance genomic medicine with AI interpretability (blog post titled 'Goodfire Announces Collaboration to Advance Genomic Medicine with AI Interpretability'). The blog excerpt is truncated in the source; full details beyond the announcement are not available.

Scale indicators14 records

Recent moves6 records

Expansion highlights7 records

Direct peers

• Weights & Biases: Enterprise ML operations and experiment-tracking platform widely used by foundation-model developers. Comparable as a developer-facing platform sold to AI labs and enterprises for training observability, though its focus is broader ML lifecycle rather than mechanistic interpretability.
• Fiddler AI: AI observability and explainability platform targeting regulated enterprises. Closely comparable to Silico in the explainability and Fortune-500 enterprise GTM motion, with explicit positioning around model monitoring and responsible AI.
• Arize AI: AI observability and evaluation platform for LLMs and ML models, providing tracing, monitoring, and drift detection in production. Directly comparable to Silico for LLMs in serving enterprise buyers who need production-grade visibility into model behavior.
• WhyLabs: AI observability platform for ML and LLM applications, including data quality, model monitoring, and drift detection. Comparable to Silico for LLMs in serving enterprise customers that need transparency into production AI behavior.
• Robust Intelligence: AI firewall and model validation platform acquired by Cisco in 2024. Comparable in offering enterprise buyers pre-deployment model testing and runtime protection, addressing adjacent aspects of the same AI-debugging pain point Goodfire targets.

Emerging players

• Conjecture: AI safety startup building tooling for alignment and interpretability of frontier models. Comparable in addressing the same customer pain points (debugging, monitoring, and aligning generative AI systems), though at an earlier commercialization stage than Goodfire.
• Apollo Research: Independent AI safety research organization focused on interpretability and evaluating frontier models for deceptive behavior. Closely comparable in mission (understanding model internals) and methodology, though it primarily publishes research rather than selling a commercial platform.

Others

• Recursion Pharmaceuticals: AI-native drug-discovery company that, like Goodfire's Life Sciences vertical, applies foundation-model interpretability and proprietary biology datasets to surface disease mechanisms. Comparable as a buyer and collaborator in the same life-sciences AI interpretability space.

Broad incumbents

• Anthropic: Frontier AI lab that maintains one of the largest in-house mechanistic interpretability research efforts (sparse autoencoders, attribution methods) and is also a Goodfire investor via its Anthology Fund. Direct overlap on interpretability R&D, but Anthropic's tooling is internal-first rather than sold as a standalone product.
• Scale AI: Large AI data and evaluation infrastructure provider serving foundation-model developers and the U.S. government. Comparable in selling into the same enterprise and frontier-AI-lab customer base for AI quality and reliability, though its core offering is data labeling and evaluation rather than mechanistic interpretability.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks7 records

Key highlights7 records

Customer concentration

• Xx.com/GoodfireAI
• LinkedInlinkedin.com/company/goodfire-ai

Compliance1 record

Revenue estimate

Valuation estimate

Number of profiles

Profiles3 records

Funding overview

Funding rounds3 records

Investors12 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Goodfire