Skip to content
Get an API key

Live · serverless · 3 regions

Llama 3.3 70B Instruct

by Meta

A 70B instruction-tuned model that handles support replies, summaries and tool calls at a mid-range price.

  • chat
  • FP8
  • dedicated
Serverless priceUSD
Input
$0.54per 1M tokens
Output
$0.72per 1M tokens

Billed per token, no minimums.

Reference

Specifications

Model ID
llama-3-3-70b
Endpoint
POST /v1/chat/completions
Publisher
Meta
Tasks
chat
Served context
128K tokens
Quantization
FP8
Decode speed
96 tok/smedian, one request
p50 first token
212 mslast 24 hours
Regions
us-east · us-west · eu-west
Dedicated
2× NVIDIA H100 per replica

Last 24 hours

Latency and speed

Median of each 15-minute step on the shared pool, measured at the API edge. Time to first token includes queueing; decode speed is for one request.

Time to first token · ms

212 msmedian

p95 230 ms

Decode speed · tok/s

96 tok/smedian

p5 91 tok/s

Show the data
Median of each 15-minute step over the last 24 hours, newest first
StepTime to first tokenDecode speed
15m ago226 ms95 tok/s
30m ago210 ms99 tok/s
45m ago210 ms98 tok/s
1h ago210 ms95 tok/s
1h 15m ago228 ms93 tok/s
1h 30m ago221 ms95 tok/s
1h 45m ago222 ms96 tok/s
2h ago215 ms94 tok/s
2h 15m ago218 ms94 tok/s
2h 30m ago230 ms91 tok/s
2h 45m ago223 ms96 tok/s
3h ago230 ms93 tok/s
3h 15m ago213 ms98 tok/s
3h 30m ago223 ms93 tok/s
3h 45m ago217 ms94 tok/s
4h ago226 ms91 tok/s
4h 15m ago317 ms74 tok/s
4h 30m ago228 ms94 tok/s
4h 45m ago209 ms96 tok/s
5h ago213 ms98 tok/s
5h 15m ago225 ms92 tok/s
5h 30m ago219 ms97 tok/s
5h 45m ago218 ms92 tok/s
6h ago215 ms94 tok/s
6h 15m ago208 ms96 tok/s
6h 30m ago213 ms93 tok/s
6h 45m ago215 ms94 tok/s
7h ago210 ms97 tok/s
7h 15m ago216 ms94 tok/s
7h 30m ago288 ms80 tok/s
7h 45m ago212 ms98 tok/s
8h ago208 ms97 tok/s
8h 15m ago215 ms94 tok/s
8h 30m ago209 ms94 tok/s
8h 45m ago220 ms96 tok/s
9h ago219 ms94 tok/s
9h 15m ago203 ms100 tok/s
9h 30m ago213 ms93 tok/s
9h 45m ago211 ms97 tok/s
10h ago203 ms98 tok/s
10h 15m ago214 ms95 tok/s
10h 30m ago214 ms93 tok/s
10h 45m ago198 ms98 tok/s
11h ago212 ms96 tok/s
11h 15m ago199 ms98 tok/s
11h 30m ago198 ms97 tok/s
11h 45m ago197 ms98 tok/s
12h ago207 ms96 tok/s
12h 15m ago207 ms97 tok/s
12h 30m ago207 ms99 tok/s
12h 45m ago213 ms98 tok/s
13h ago206 ms98 tok/s
13h 15m ago200 ms101 tok/s
13h 30m ago200 ms96 tok/s
13h 45m ago196 ms99 tok/s
14h ago197 ms98 tok/s
14h 15m ago211 ms96 tok/s
14h 30m ago198 ms96 tok/s
14h 45m ago197 ms100 tok/s
15h ago197 ms96 tok/s
15h 15m ago206 ms98 tok/s
15h 30m ago205 ms98 tok/s
15h 45m ago211 ms99 tok/s
16h ago288 ms76 tok/s
16h 15m ago210 ms97 tok/s
16h 30m ago199 ms97 tok/s
16h 45m ago206 ms97 tok/s
17h ago196 ms99 tok/s
17h 15m ago212 ms97 tok/s
17h 30m ago205 ms95 tok/s
17h 45m ago195 ms102 tok/s
18h ago203 ms100 tok/s
18h 15m ago205 ms99 tok/s
18h 30m ago204 ms96 tok/s
18h 45m ago199 ms101 tok/s
19h ago197 ms97 tok/s
19h 15m ago205 ms98 tok/s
19h 30m ago214 ms98 tok/s
19h 45m ago213 ms93 tok/s
20h ago214 ms95 tok/s
20h 15m ago204 ms99 tok/s
20h 30m ago217 ms93 tok/s
20h 45m ago208 ms97 tok/s
21h ago214 ms93 tok/s
21h 15m ago304 ms76 tok/s
21h 30m ago221 ms93 tok/s
21h 45m ago212 ms98 tok/s
22h ago213 ms98 tok/s
22h 15m ago219 ms95 tok/s
22h 30m ago219 ms95 tok/s
22h 45m ago226 ms94 tok/s
23h ago210 ms96 tok/s
23h 15m ago214 ms93 tok/s
23h 30m ago222 ms94 tok/s
23h 45m ago223 ms95 tok/s
24h ago215 ms95 tok/s

API

Make a request

The same request shape as every other model on /v1/chat/completions. Switching between them is a change to one string.

curl https://api.weir.example/v1/chat/completions \
  -H "Authorization: Bearer $WEIR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3-3-70b",
    "messages": [
      { "role": "user", "content": "What does a KV cache do?" }
    ]
  }'
POST /v1/chat/completionsapi.weir.example

Dedicated

Run it on GPUs reserved for you

Replica
2× H100
Replica rate
$4.98 / hour
Peak throughput
6,000 tok/s
Break-even, 1 replica
5.77B tok/mo

Replicas reserved for you, billed by the second, with a 99.95% monthly uptime SLA. Break-even is where one replica costs the same as serverless, at half input and half output tokens.

Run on dedicated GPUsCompare costs

Compare

Similar models