API model string
open-mixtral-8x22bContext window
65K tokens
Max output not published
- Knowledge cutoff
- not published
- Released
- 2024
- Modalities
- text in, text out
Last verified July 10, 2026
Rate limits by tier
Mixtral 8x22B is a legacy model with an unconfirmed first-party API status on La Plateforme as of July 2026. Realistic access paths are third-party hosted inference (Together AI, Fireworks, DeepInfra) or self-hosting the Apache 2.0 open weights. Limits below reflect the actual access paths available.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| La Plateforme (first-party, historical) | La Plateforme account — current availability not confirmed; verify on mistral.ai/pricing | not published (legacy) | not published | not published | — | Mixtral 8x22B was historically listed on La Plateforme at $2/$6 per MTok. Current first-party status is unconfirmed as of July 2026. Do not rely on this endpoint without verifying availability on mistral.ai/pricing. |
| Third-party hosted (Together AI, Fireworks, DeepInfra) | Account on the respective host platform | host-dependent (check each host's dashboard) | host-dependent | host-dependent | — | Open weights mean multiple hosts offer Mixtral 8x22B inference. Rates, limits, and reliability vary per host. Compare pricing on each host's pricing page before committing. |
| Self-hosted (Apache 2.0) | ~141GB VRAM at FP16 (e.g., 8× A100 80GB); ~70GB at int4 (quantized) | unlimited (limited by your hardware) | unlimited | unlimited | — | Apache 2.0 license permits commercial self-hosting at zero per-token cost. MoE architecture activates only ~39B of 141B total parameters per token, giving good performance-per-FLOP for self-hosted inference. |
Swipe the table sideways to see every limit column.
- 1.Mixtral 8x22B open weights are Apache 2.0 — commercial use is permitted with no licensing fee.
- 2.Third-party host prices change frequently; verify on the specific host's pricing page before quoting.
- 3.La Plateforme first-party availability is unconfirmed as of July 2026 — the model is no longer a promoted SKU on the main pricing page.
Limits verified against the Mistral AI docs, July 10, 2026.
Token pricing
What you pay per million tokens (USD). Input and output are billed separately.
Input
$2.00
per 1M tokens
Output
$6.00
per 1M tokens
- The $2.00/$6.00 per MTok figures reflect the historical La Plateforme rate card for Mixtral 8x22B — current first-party availability is unconfirmed. Treat these as historical reference only and verify on mistral.ai/pricing.
- Third-party host pricing for Mixtral 8x22B varies by host and is not captured in this brief — verify on Together AI, Fireworks, or DeepInfra pricing pages directly.
- Self-hosting via Apache 2.0 open weights = $0 per-token cost (infrastructure only).
- Batch discount availability on this legacy endpoint is not confirmed — verify if La Plateforme endpoint is still active.
Run your own numbers
Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.
Estimated open-mixtral-8x22b spend
$90/mo
Input: $60
Output: $30
30M in × $2.00 + 5M out × $6.00 = $90
Same volume, priced across models
- Llama 4 ScoutCheapest$3.90
- Mistral Medium 3.5$22
- open-mixtral-8x22bThis model$90
- Mistral Large 3$90
Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to open-mixtral-8x22b only. Estimates for comparison; real bills vary with request shape and long-context surcharges.
open-mixtral-8x22b vs the alternatives
Mixtral 8x22B is a legacy MoE model; this comparison helps you decide whether to self-host it or migrate to an actively developed alternative.
| Aspect | open-mixtral-8x22b | Mistral Medium 3.5 | Mistral Large 3 | Llama 4 Scout |
|---|---|---|---|---|
| Architecture | MoE sparse (39B active / 141B total params) | Dense | Dense | MoE sparse |
| Context window | 65K | 128K | 128K | 10M native (host-capped) |
| First-party API | Unconfirmed (legacy) | Active (La Plateforme) | Active (La Plateforme) | Third-party only |
| Input $/MTok | $2.00 historical (verify) | $0.40–1.50 (verify) | $2.00 (verify) | $0.08–0.15 (host) |
| Output $/MTok | $6.00 historical (verify) | $2.00–7.50 (verify) | $6.00 (verify) | $0.30–0.60 (host) |
| Open weights | Yes (Apache 2.0) | No | No | Yes (Llama 4 license) |
| EU residency (first-party) | Unconfirmed | Yes (Paris) | Yes (Paris) | No |
| Active development | No (legacy) | Yes | Yes | Yes |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
HTTP 429 Too Many Requestsrate_limit_errorRetry-After: <seconds>Why it happens & how to fix it
La Plateforme endpoint inactive for legacy model
Confirm first-party availability on mistral.ai/pricing. If the endpoint is inactive, switch to a third-party host (Together AI, Fireworks, DeepInfra) or self-host the open weights. Do not build a production dependency on an unconfirmed legacy endpoint.
Third-party host RPM limit hit
Implement exponential backoff honoring the host's Retry-After header. Each third-party host has different rate limits — check their specific dashboards. Consider distributing load across multiple hosts if one is consistently rate-limiting.
Self-host GPU out-of-memory (OOM)
Mixtral 8x22B requires ~141GB VRAM at FP16 (8+ A100 80GB). Use a quantized checkpoint (int4) to reduce VRAM to ~70GB. Consider switching to a smaller model if hardware is the constraint.
Context length exceeded (65K tokens)
Mixtral 8x22B has a 65K context window — narrower than current models. For longer contexts, switch to Mistral Large 3 (128K) or Mistral Medium 3.5 (128K). Truncate input or use chunking strategies for the legacy endpoint.
Model string not found on host
Each host uses a different model string for Mixtral 8x22B. La Plateforme (if active): 'open-mixtral-8x22b'. Together AI: 'mistralai/Mixtral-8x22B-Instruct-v0.1'. Verify the exact model string in your chosen host's documentation.
Retry strategy
Standard HTTP 429 with Retry-After header — honor the header wait time precisely. For exponential backoff fallback: delay = min(2^attempt, 60) seconds ± 10% jitter. Behavior depends on which host you use. For La Plateforme (if active): OpenAI-compatible base URL https://api.mistral.ai/v1. For third-party hosts, use their specific base URLs.
1import OpenAI from 'openai';23// Update BASE_URL and MODEL_STRING for your chosen host:4// La Plateforme (if active): https://api.mistral.ai/v1 / 'open-mixtral-8x22b'5// Together AI: https://api.together.xyz/v1 / 'mistralai/Mixtral-8x22B-Instruct-v0.1'6// Fireworks: https://api.fireworks.ai/inference/v1 / 'accounts/fireworks/models/mixtral-8x22b-instruct'7const BASE_URL = 'https://api.mistral.ai/v1';8const MODEL_STRING = 'open-mixtral-8x22b'; // verify current string with your host910const client = new OpenAI({11 apiKey: process.env.MISTRAL_API_KEY, // use host-specific env var12 baseURL: BASE_URL,13});1415async function callMixtral8x22bWithRetry(16 prompt: string,17 maxRetries = 518): Promise<string> {19 for (let attempt = 0; attempt < maxRetries; attempt++) {20 try {21 const response = await client.chat.completions.create({22 model: MODEL_STRING,23 messages: [{ role: 'user', content: prompt }],24 });25 return response.choices[0].message.content ?? '';26 } catch (err: unknown) {27 const error = err as { status?: number; headers?: Record<string, string>; message?: string };28 if (error.status === 429) {29 const retryAfter = parseFloat(error.headers?.['retry-after'] ?? '0');30 const jitter = 1 + (Math.random() * 0.2 - 0.1);31 const delay = retryAfter > 032 ? retryAfter * 100033 : Math.min(Math.pow(2, attempt) * 1000 * jitter, 60000);34 console.warn(`Rate limited. Retrying in ${(delay / 1000).toFixed(1)}s (attempt ${attempt + 1}/${maxRetries})`);35 await new Promise(resolve => setTimeout(resolve, delay));36 } else {37 throw err;38 }39 }40 }41 throw new Error('Max retries exceeded for Mixtral 8x22B API call');42}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Self-host (Apache 2.0 open weights)
Time to set up GPU infrastructureDownload open weights from Hugging Face (apache-2.0 license). Deploy on ~141GB VRAM FP16 or ~70GB VRAM with int4 quantization.
Unlocks: Zero per-token cost, unlimited throughput limited only by hardware, full data control
Third-party host (free/trial tier)
Immediate upon account creationSign up on Together AI, Fireworks, or DeepInfra. Each offers a free or trial tier for Mixtral 8x22B inference.
Unlocks: Quick access to hosted Mixtral 8x22B inference without GPU infrastructure
Third-party host (paid)
Immediate upon plan upgradeUpgrade to a paid plan on your chosen host. Each has their own tier structure — verify limits on the host's dashboard.
Unlocks: Higher RPM/TPM limits, SLA, dedicated support (host-dependent)
La Plateforme (if active)
Verify availability first on mistral.ai/pricingCheck Admin Console → Limits on La Plateforme if you need first-party Mistral infrastructure. Contact sales@mistral.ai for enterprise usage if the endpoint is confirmed active.
Unlocks: EU Paris data residency, SAML SSO, audit logs (enterprise tier)
Cut your token spend
Apache 2.0 self-hosting for zero per-token cost
-100% per-token cost (infrastructure only)The open weights permit commercial use without a license fee. For high-volume workloads where data privacy is paramount, self-hosting on your own GPU infrastructure eliminates per-token cost entirely.
Quantized inference (int4)
~50% reduction in VRAM requirement with acceptable quality lossRun Mixtral 8x22B at int4 quantization to reduce VRAM from ~141GB (FP16) to ~70GB. This cuts hardware cost roughly in half for self-hosted deployments with acceptable quality tradeoff for most tasks.
Migrate to Mistral Small/Medium 3.x
Longer context (128K vs 65K), active development, potentially lower hosted costFor new projects, Mistral Medium 3.5 is the recommended migration path: it offers 128K context (vs 65K), active development, EU data residency via La Plateforme, and a confirmed first-party endpoint. Migration is the primary recommendation for any new workload.
Use MoE sparse architecture advantage
Good performance-per-FLOP for self-hosted inferenceMixtral 8x22B activates only ~39B of its 141B total parameters per token. This makes self-hosted inference more efficient per token than a dense 141B model would be, improving throughput on the same hardware.
Compare third-party host prices before committing
Significant cost variation across hosts (can differ meaningfully per MTok)Together AI, Fireworks, DeepInfra, and others may price Mixtral 8x22B differently. Check each host's pricing page before committing to a contract — rates vary and change over time.
Frequently asked questions
Does Mixtral 8x22B still have a working API?
As of July 2026, Mixtral 8x22B's first-party availability on Mistral's La Plateforme is unconfirmed — it is no longer a promoted SKU on the main pricing page. You can still access it through third-party inference hosts (Together AI, Fireworks, DeepInfra) or by self-hosting the Apache 2.0 open weights. Verify current first-party status on mistral.ai/pricing before building any dependency on the La Plateforme endpoint.
What is Mixtral 8x22B pricing?
The historical La Plateforme rate was $2.00 per million input tokens and $6.00 per million output tokens — treat these as historical reference only since current first-party availability is unconfirmed. Third-party host pricing varies; verify on Together AI, Fireworks, or DeepInfra pricing pages directly. Self-hosting via the Apache 2.0 open weights costs $0 per token (infrastructure only).
How do I self-host Mixtral 8x22B?
Download the open weights from Hugging Face (Apache 2.0 license permits commercial use at no fee). You need approximately 141GB VRAM at FP16 — for example, 8× A100 80GB GPUs — or roughly 70GB VRAM using int4 quantization. The MoE architecture activates only ~39B of 141B total parameters per forward pass, giving reasonable throughput relative to the total parameter count.
What should I migrate to from Mixtral 8x22B?
Mistral Medium 3.5 is the recommended replacement: it has 128K context (vs Mixtral 8x22B's 65K), active development, a confirmed first-party endpoint on La Plateforme with EU data residency, and comparable pricing. For higher quality, Mistral Large 3 is the flagship alternative. Both are actively developed and have confirmed La Plateforme endpoints. RapidDev can help scope a migration plan for your existing Mixtral 8x22B integration — rapidevelopers.com/contact.
Is Mixtral 8x22B open source?
Yes — Mixtral 8x22B is released under the Apache 2.0 license, which permits commercial use, modification, and distribution without a licensing fee. The open weights are available on Hugging Face. This makes it one of the more permissively licensed large-scale MoE models available for self-hosting.
How much VRAM does Mixtral 8x22B need?
Approximately 141GB VRAM at FP16 precision — equivalent to 8× A100 80GB GPUs or similar high-VRAM setup. With int4 quantization, the requirement drops to approximately 70GB VRAM (roughly 2× A100 80GB or 4× A100 40GB), with an acceptable quality tradeoff for most use cases. The MoE architecture activates only ~39B of the 141B total parameters per token, so inference throughput is better than a dense 141B model on the same hardware.
What is the Mixtral 8x22B context window?
Mixtral 8x22B supports a 65K token context window. This is narrower than current actively developed models — Mistral Medium 3.5 and Mistral Large 3 both offer 128K tokens. If your use case requires longer contexts, this is a primary reason to migrate to a current Mistral model.
How do I fix HTTP 429 errors on Mixtral 8x22B?
The fix depends on your access path. For third-party hosts (Together AI, Fireworks, DeepInfra): honor the Retry-After header and use exponential backoff — delay = min(2^attempt, 60) seconds ± 10% jitter. Each host has different rate limits; check their dashboards for your specific ceilings. For La Plateforme (if active): same OpenAI-compatible retry pattern with base_url='https://api.mistral.ai/v1'. For self-hosted deployments, 429 is not applicable — bottleneck will be GPU throughput instead.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.