The Open-Source Strategy

Why open-source models

Together's strategic bet was that customers would care about more than absolute frontier capability:

• Cost. Per-token pricing on optimized open-source serving can be 50-80% cheaper than equivalent OpenAI / Anthropic API pricing.
• Customization. Open-source models can be fine-tuned. Closed-source frontier models can't be fine-tuned at the level open-source allows.
• Data control. Some customers don't want to send sensitive data through proprietary model APIs.
• Regulatory. Some sectors require model traceability that open-source enables.
• Optionality. Customers want to avoid lock-in to a single closed-source provider.

These motivations support a real market for open-source-hosted inference. Together's bet was that the market would be large enough to support a platform business.

The curation problem

Open-source model releases happen constantly. Llama family alone has multiple variants per release; Mistral, Qwen, DeepSeek release at their own cadence. Customers don't want to evaluate every model release.

Together's curation:

• Maintain a catalog of the genuinely competitive open-source models.
• Deprecate models that have been surpassed.
• Publish benchmarks and quality comparisons.
• Maintain instruction-tuned variants for chat-style use cases.
• Surface the right model for the right use case.

This is a real product investment that customers value. Doing this well requires deep model evaluation expertise — exactly what Together's research lineage supports.

Performance gap

How does open-source quality compare to closed-source frontier?

• 2022-early-2023: large gap. GPT-4 was meaningfully ahead of any open-source model.
• Late 2023-2024: gap narrowed. Llama 2/3, Mixtral, etc. were competitive on many benchmarks.
• 2024-2026: gap variable by task. For many tasks, open-source is close to or matches closed-source frontier. For some (reasoning-heavy, coding, agentic), closed-source retains a lead.

The narrowing gap is the strongest tailwind for Together's strategy. The remaining gap is the constraint.

Commercial advantages

From a customer's perspective, the per-token economics on open-source-hosted inference at Together vs closed-source frontier APIs:

• Often 50-80% cheaper per token.
• Slightly different latency profile (variable depending on model and serving optimization).
• Can be much cheaper for high-volume workloads where the latency advantage of closed-source frontier isn't needed.

For workloads where open-source quality is "good enough" — many enterprise use cases — the cost savings make Together a strong economic choice.

Ecosystem positioning

Together has positioned itself as a friendly counterparty to the open-source ecosystem rather than a competitor:

• Contributing back to open-source projects (FlashAttention, etc.).
• Releasing models (RedPajama, etc.).
• Publishing research that benefits the broader community.
• Hosting models from many providers without playing favorites.

This posture builds goodwill that translates into customer trust. Compared to managed-inference competitors who are more closed about their stack, Together's research-and-open-source presence is differentiated.

Risks

• Closed-source frontier reasserts dominance. If OpenAI / Anthropic / Google maintain enough quality lead, customers stick with closed APIs.
• Open-source quality plateaus. Without continued investment from Meta and others, the catalog of competitive models could stagnate.
• Hyperscalers absorb open-source. AWS Bedrock, Azure AI, GCP Vertex all offer open-source model hosting. Customers may default to their existing cloud rather than to Together.
• Margin compression. As open-source-inference becomes commoditized, prices fall and Together's margins compress.

Takeaway

The open-source strategy is the load-bearing strategic bet at Together. It has aged well so far; whether it continues to age well depends on the open-source ecosystem maintaining momentum and on Together capturing share faster than hyperscalers can absorb the same workloads. The next chapter looks at the research credibility that supports the strategy.