Llama 3.3 70B

“The dependable middle option. Not the smartest, not the cheapest, but the most operationally proven open model I can standardize on for text.”

For a buyer, 3.3 70B is the low-risk open default. Eighteen-plus months of provider experience means chat templates, fine-tuning recipes, quantization paths, and edge cases are all well-documented, and it carries the broadest tool ecosystem of any Llama. It deploys cleanly on a single node and is supported on every major cloud. The trade-offs are the December 2023 cutoff, the lack of vision, and that it is now superseded by Llama 4 Scout for new builds. For text-only workloads on a 24-month horizon where operational maturity is the priority, it remains a defensible standard.

“Its moat is maturity, not capability. In a market sprinting on benchmarks, 'boring and proven' is a smaller but real square.”

Strategically, 3.3 70B occupies the "proven open text workhorse" position. It does not lead any benchmark in 2026 and is outflanked by its own successor (Scout) on context and vision and by newer open models (Qwen 3, DeepSeek) on reasoning. Its differentiation is purely operational maturity and instruction-following reliability. Market timing has passed its peak — the open-weights conversation has moved to MoE and long context — so its strategic relevance is shrinking even as its installed base stays large. A durable present, a fading future.

“Excellent TCO but no longer best-in-class — Scout undercuts it at almost every provider, and dense 70B needs more GPUs for the same throughput.”

The economics are strong but the leadership has passed. DeepInfra runs it at ~$0.10/$0.30 blended; Groq at $0.59/$0.79; Bedrock at $0.72–$0.90. Above ~500M tokens/month, self-hosting on reserved H100s beats managed by 3–5x. The catch is the dense architecture: 70B needs roughly 4x the GPUs of a 17B-active MoE like Scout for equivalent throughput, so if you are sizing fresh, Scout usually wins the $/throughput math. If you already run a 70B fleet, it stays cheap. Predictable, well-understood, but no longer the value frontier.

“The most boring, predictable Llama in production — and that's a compliment. Stable template, forgiving fine-tunes, consistent across hosts.”

Builders get the most predictable Llama in production. The chat template is stable across Together, Fireworks, Groq, and Bedrock; function-calling formats are well-documented; behavior is consistent host-to-host (a real contrast to the Llama 4 template drift). Fine-tuning is fast and forgiving, and LoRA adapters generalize well. The 128K context covers ~95% of real workloads. No vision means no accidental image-handling surprises. The downside for builders is purely capability: it is text-only, dense, and behind the frontier — but for shipping reliable text features it is hard to beat.

“Fluent, reliable, slightly conservative — correct and on-format, rarely surprising. Great backstage, underwhelming as a personality.”

End users get a fluent, reliable, somewhat conservative chat partner. Refusal rates are sensible, instruction-following is the best in the open tier, and latency on Groq/Cerebras is sub-second. The feel is competent but lacks the warmth of Claude or the wit of GPT-5 — answers are correct and on-format but rarely delightful. For embedded SaaS assistants, support backends, and any context wanting predictable on-brief output, it is exactly right. For a flagship consumer chatbot where personality is the product, users feel underwhelmed.

“A genuinely good text model whose marketing aged honestly — but it's text-only, Instruct-only, and its December 2023 brain shows on anything recent.”

Adversarially, 3.3 70B is refreshingly honest — its benchmark claims (IFEval 92.1, MATH 77.0) hold up and there is no LMArena experimental-checkpoint shenanigans like Llama 4. The real weaknesses are structural, not deceptive: text-only with no vision, Instruct-only with no base checkpoint, a dense architecture that loses on cost-per-throughput to MoE, and a December 2023 cutoff that surfaces on recent libraries, events, and product names. It is also now superseded by Meta's own Scout for most new builds. The honest verdict: a very good 2024 text model that remains useful but is no longer the open frontier.