Migrating off Llama 3.3 70B
Llama 3.3 70B (Groq) — retired 2026-08-16.
Source · Verified 2026-08-14.
Groq is decommissioning Llama 3.3 70B on its LPU inference platform. This isn't a Meta model deprecation — Llama 3.3 70B's weights aren't going anywhere — it's Groq specifically pulling the model from its hosted API, which is the only place we route requests for it.
What should I use instead of Llama 3.3 70B?
Llama 3.3 70B, from Groq, was retired 2026-08-16. There is no direct 1:1 successor — see the cross-provider alternatives below for the closest match.
What actually changes
| Llama 3.3 70B | — |
|---|
| Blended $/M | — | — |
| Context window | 128,000 tokens | — |
| Max output | 32,768 tokens | — |
| Modalities | text | — |
| Tokens / sec | unchanged / not measured | not measured |
Pricing from our live pricing registry; speed from our benchmark leaderboard where measured.
Gotchas
- There is no first-party Groq successor for this exact model at shutdown time. Meta's own Model API (Muse Spark 1.3) is the intended long-term replacement, now selectable through our gateway.
- Llama 3.3 70B has no native reasoning/thinking mode. If you route to a reasoning-capable replacement (GPT-OSS 120B, DeepSeek V4 Pro), expect a different response shape if you were parsing raw completions rather than using our unified response format.
- Groq is known for aggressive rate limits during high-demand windows relative to token-metered providers — moving to a token-billed model (rather than another Groq-hosted one) changes your cost predictability at scale, not just your per-token price.
- If you were relying on Groq's sub-100ms time-to-first-token for a latency-sensitive path, GPT-OSS 120B on Groq is the closest speed match; moving off Groq entirely will very likely be slower even if it's cheaper.
Alternatives beyond the obvious successor
Also Groq-hosted — keeps the same LPU latency profile, adds reasoning mode.
Groq-hosted multimodal alternative with vision support Llama 3.3 70B lacked.
Off Groq entirely — token-billed, no LPU rate-limit exposure.
FAQ
Is Llama 3.3 70B being shut down?
Yes — it was retired on 2026-08-16. Source: https://console.groq.com/docs/deprecations.
What should I switch to from Llama 3.3 70B?
There is no direct 1:1 successor — see the alternatives section for cross-provider options.
Will switching cost more or less?
See the pricing comparison table below for exact numbers.
Batch 50 · llama-3-3-70b decision and evidence contributions. Surface verification: 2026-08-14. These are route-local, server-rendered fixtures; unavailable values are not inferred.
Groq-endpoint identity and shutdown resolver
Frozen Batch 50 fixture board. Formula / decision rule: resolved = provider + host + exact endpoint ID + shutdown evidence + successor endpoint Boundary: A Meta-weight retirement never transfers shutdown evidence to other serving hosts.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-3-3-70b-m1-r1
Groq llama-3.3-70b-versatile · official notice | host=api.groq.com; endpoint=llama-3.3-70b-versatile; notice=groq-deprecation-policy; shutdown=Unavailable per public docs 2026-08-14
Groq notice confirms endpoint is deprecated but published shutdown date is not found in accessible first-party documentation. | shutdown date = Unavailable; do not invent; check Groq console | UNAVAILABLE — date join required. |
batch50-llama-3-3-70b-m1-r2
Third-party proxy · same endpoint string | host=proxy.example; endpoint=llama-3.3-70b-versatile; weights=meta; notice=none; shutdown=copied from groq
The third-party host carries the same string but has no Groq lifecycle authority. | shutdown evidence does not transfer; host identity must rejoin | FAIL CLOSED — host identity mismatch. |
batch50-llama-3-3-70b-m1-r3
Other OSS hoster · Meta weights · independent lifecycle | host=other-oss-hoster; endpoint=meta-llama-3.3-70b; weights=meta; groq-notice=not-applicable
The Meta weights are intact on another host with its own lifecycle; Groq deprecation is irrelevant here. | lifecycle state = independent; groq shutdown does not apply | PASS WITH SEPARATION — host-qualified only. |
Provenance: Batch 50 llama-3-3-70b module 1 first-party evidence, surface verification date 2026-08-14. Groq model deprecation notices. Missing joins fail closed.
Workload-dependency inventory and migration schedule
Frozen Batch 50 fixture board. Formula / decision rule: migration start = latest(dependency discovery deadline, successor availability confirmation) Boundary: Successor capability must be verified per workload before traffic is cut over.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-3-3-70b-m2-r1
Real-time chat · batch summarisation · tool-use pipeline | workload=chat/batch/tool; endpoint=llama-3.3-70b-versatile; discovered=2026-08-14; tool support=verify per successor
Three workload classes have different sensitivity to latency, throughput, and tool support. | migration order: verify tool support → batch → real-time | SCHEDULED — workload priority set. |
batch50-llama-3-3-70b-m2-r2
Context-window-dependent workload · 128K prompt | workload=long-context; prompt=128000; endpoint=llama-3.3-70b-versatile; context=128K; successor-context=Unavailable until verified
Successor context ceiling must be joined before cut-over is approved for long-context workloads. | migration=blocked until successor context join; keep canary on old endpoint | BLOCKED — capability join required. |
batch50-llama-3-3-70b-m2-r3
Canary replay · rollback reserve | old=llama-3.3-70b-versatile; canary share=5%; rollback window=7d; semantic match=measured; judge=human
A 5% canary with a measured semantic match and 7-day rollback window is the minimum safe gate. | promote if semantic delta < threshold over 7 days; revert immediately if quality drops | GATE — canary + rollback required. |
Provenance: Batch 50 llama-3-3-70b module 2 first-party evidence, surface verification date 2026-08-14. Groq model deprecation notices. Missing joins fail closed.
Groq request-envelope compatibility receipt
Frozen Batch 50 fixture board. Formula / decision rule: compatible = endpoint in active catalog AND all used parameters documented AND response schema unchanged Boundary: Parameter support may differ between Groq and the successor even when both serve Llama weights.
| Frozen fixture / field ID | Joined inputs and observation | Calculated result | State |
|---|
batch50-llama-3-3-70b-m3-r1
stop sequence · presence penalty · JSON mode | endpoint=llama-3.3-70b-versatile; params=stop/presence_penalty/response_format; groq-support=documented; successor-support=Unavailable until rechecked
Parameter support must be rechecked per candidate successor endpoint on Groq. | compatibility=Unresolved until successor param docs joined | UNRESOLVED — param join required. |
batch50-llama-3-3-70b-m3-r2
Streaming · tool calls · system prompt | endpoint=llama-3.3-70b-versatile; streaming=yes; tools=yes; system=yes; successor=candidate; schema delta=check usage field
Usage field and tool call format may differ; streaming chunk order should be replayed. | surface schema delta in canary before promoting | VERIFY — schema replay gate. |
batch50-llama-3-3-70b-m3-r3
Hard-coded endpoint string · dynamic model selector | pattern=hard-coded; endpoint=llama-3.3-70b-versatile; env var=no; rollback=impossible without redeploy
Hard-coded strings require a full redeploy; dynamic selectors can be hot-patched. | convert to env var before shutdown; rollback then requires env update only | ACTION REQUIRED — refactor first. |
Provenance: Batch 50 llama-3-3-70b module 3 first-party evidence, surface verification date 2026-08-14. Groq model deprecation notices. Missing joins fail closed.
Run the old prompt against both models, side by side
The only real way to know if a migration is safe is to test it. One workspace, both models, your actual prompts.
Try It Free