Llama 5

What it is

Llama 5 is Meta's flagship open-weights model: the broad, well-supported base that the note on this page positions for fine-tuning and self-hosting. The Llama line has been the default starting point for open model work since 2023, and the ecosystem around it — quantisations, serving stacks, fine-tuning recipes, third-party hosts — is the practical reason teams pick it.

Verify the licence before assuming "open" means what it does for genuinely permissive models: Llama licences have always been custom terms with acceptable-use conditions attached. The exact window and parameter count for Llama 5 are in Meta's model card, not on this board.

How the composite is built

Overall 84 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. Llama 5 is the flattest profile in its band — 77 to 86 across all six axes, with no weak point and no standout.

Coding 86

Six behind DeepSeek V4's 92 and two behind Terra's 88. Competent across languages; the specialist code models still beat it.

Terminal 83

Between Terra's 84 and Mistral Large 3's 81, and twelve behind Opus 5's 95. Supervised loops are fine; hands-off operation is not the design centre.

Reasoning 86

Level with Kimi k3 and GLM-5, seven behind DeepSeek R2's 93. Solid planning for an open generalist.

Tool use 84

One behind Nemo 3 Ultra's 85 and six behind Sonnet 5's 90. Reliable enough for structured calling with validation in the harness.

Context 84

Level with Qwen 4 Coder and Nemo 3 Ultra. Module-scale work, not whole-repository reads.

Speed 77

One behind Qwen 4 Coder's 78 on hosted inference — but with no gateway listed on this board, this read is less meaningful than for hosted models. Self-hosted, the number is your hardware.

Where it fits

As a base model: fine-tuning for a domain, self-hosted deployments where data residency is fixed, and any project that wants the deepest tooling ecosystem in open models. It is the sensible default when you need one open model that does everything reasonably and you intend to adapt it.

Limits

• No price or index is published here, so Value is Unrated — the cost of running Llama 5 is a hardware question, and GPU hours are the real budget line.
• It leads nothing. Every axis is within a few points of the open-weights pack; if one capability decides the project, a specialist beats it.
• The licence is not a permissive open-source licence. Read the terms — commercial redistribution and acceptable-use clauses have bitten adopters before.
• No gateway listing on this board. Hosted Llama inference is available from many third parties; those rates are not measured here.

Price and access

No published price on this board. Access is the weights from Meta's channels, then your own GPUs or a third-party host. Meta's own hosted listings appear elsewhere on the board unscored — the Muse Spark entries at $1.25/$4.25 — and are worth checking if you want Meta models without self-hosting. Last scored 15 Sep 2026.

Alternatives on this board

• DeepSeek V4 — 86 at $0.78/$1.57 with Coding 92 and open weights. Two points up with a known hosted rate.
• Qwen 4 Coder — 84 with Coding 90, the open-weights rival if code is the main workload.
• GLM-5 — 82 at $0.60/$1.92 with three gateways, if hosted cost matters more than ecosystem breadth.
• Llama 5 Scout — 81 with Speed 85, the smaller sibling for high-throughput serving.

Sources