Ask AI

Fireworks.ai

Fireworks.ai provides production inference for open-weight models. AvalAI exposes selected Fireworks.ai-hosted models through compatible API routes while the model owner remains the original developer, such as Meta or NVIDIA.

Available Models

muse-spark-open — Meta's open-weight 816B/42B-active model, available from Tier 0
muse-glimmer-30b — Meta's 30B multimodal agentic model
nemotron-3.5-lightning — NVIDIA's efficient 30B/3B-active reasoning model
nemotron-3-ultra — NVIDIA's large-scale reasoning and agentic model

API Endpoint Support

Model	`v1/chat/completions`	`v1/messages`	`v1/responses`
`muse-glimmer-30b`	✅ Full	✅ Full	⚠️ Partial
`nemotron-3.5-lightning`	✅ Full	✅ Full	⚠️ Partial
`nemotron-3-ultra`	✅ Full	✅ Full	⚠️ Partial

muse-spark-open uses the Chat Completions baseline documented below; this release does not extend the existing endpoint matrix to it.

Partial Responses support means some Responses-only fields or stateful features may not be available. Validate the exact request shape before production rollout.

Key Features

Capabilities vary by model and route; use each model's own support section before choosing optional features.

OpenAI-compatible Chat Completions access
Anthropic-compatible Messages access
Function calling and agentic workflows
Provider-side implicit prompt caching where eligible
Cached-input pricing for repeated prompt prefixes

Muse Spark Open

Muse Spark Open, added to AvalAI on October 25, 2026, is developed by Meta and routed through Fireworks.ai. Meta's official overview describes an open-weight sparse mixture-of-experts model with 816B total parameters and 42B active parameters per token, perception encoders, and text output. Meta's deployment guide lists Fireworks among its hosted providers. The catalog's owned_by: nvidia value is routing metadata; it is not evidence that NVIDIA developed the model.

Access, Limits, and Choosing the Route

Use muse-spark-open as the exact AvalAI request ID. The upstream Fireworks ID is accounts/fireworks/models/muse-spark-open; do not put that provider string, a Fireworks key, or a Fireworks URL into the AvalAI example. Access starts at Tier 0. AvalAI records 1,048,576 maximum input tokens but no output-token ceiling for this route. The input ceiling is not an output allowance or a guaranteed combined request budget. Check the exact model's account limits and evaluate larger requests independently.

Chat Completions is the documented AvalAI text baseline for this addition. The current catalog records reasoning, vision, and function calling. Meta describes text/image input, tool calling, automatic caching, and OpenAI-compatible serving on Fireworks upstream; optional payloads still need AvalAI checks. Upstream Messages support, self-hosted video/audio examples, JSON mode, reasoning controls, computer use, and dedicated/private deployments do not establish those features on this AvalAI route. This addition qualifies no new Messages or Responses support badge. Start without optional reasoning or sampling parameters, then test the exact feature before depending on it.

Choose Spark Open when you want Meta's open-weight model through hosted inference without operating GPUs, or need Tier 0 access. Choose direct Muse Wonder when its software-engineering and long-horizon-agent design fits the task and Tier 1 access is available; it has different routing and fee treatment. For smaller workloads, compare the existing Glimmer and Nemotron options below. Evaluate answer quality, time to first answer, output-token use, and tool correctness on your own workload rather than treating a provider benchmark as an AvalAI guarantee. Your application executes tools and reviews proposed actions.

AvalAI Pricing and Prompt Caching

All amounts below are USD per 1 million tokens and are current AvalAI billed rates:

Charge	AvalAI billed rate
Uncached input	$1.25
Cached input	$0.15
Output	$4.25

No new processing fee applies to this Fireworks.ai route. Do not apply the direct Meta 15% exception to Spark Open or multiply the current catalog rates again. These are the standard route's prices; upstream priority modes, dedicated capacity, and separate provider plans do not change the example's AvalAI contract.

Meta's Fireworks deployment documentation describes automatic prompt caching. Keep stable instructions, examples, and reusable history at the start and changing content at the end. When returned, usage.prompt_tokens_details.cached_tokens counts cached tokens within prompt_tokens, not extra tokens. For P total input tokens, C cached input tokens, and O output tokens, the token estimate is ((P - C) × 1.25 + C × 0.15 + O × 4.25) / 1,000,000 USD. A cached-input price does not guarantee a cache hit, retention duration, cache-control parameter, or write charge; use actual usage fields and prompt-caching guidance. The example does not set upstream caching controls.

Bounded Chat Completions Example

Keep AVALAI_API_KEY server-side. Each program sends one request to POST https://api.avalai.ir/v1/chat/completions, uses the OpenAI SDK base URL https://api.avalai.ir/v1, caps generation at max_tokens: 256, sets a 60-second request timeout, and disables automatic retries. curl also limits connection time to 10 seconds. This is a small compatibility example. It does not validate long reasoning runs or set a maximum output capability for the model. The upstream model can spend its token budget on reasoning; inspect the actual answer and finish_reason rather than assuming every successful request contains finished text.

BashPythonJavaScript

bash

curl --fail-with-body --silent --show-error \
  --connect-timeout 10 --max-time 60 --retry 0 \
  https://api.avalai.ir/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -d '{
    "model": "muse-spark-open",
    "messages": [{"role": "user", "content": "Give a concise checklist for reviewing an API change."}],
    "max_tokens": 256
  }'

python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AVALAI_API_KEY"],
    base_url="https://api.avalai.ir/v1",
    timeout=60.0,
    max_retries=0,
)
response = client.chat.completions.create(
    model="muse-spark-open",
    messages=[
        {"role": "user", "content": "Give a concise checklist for reviewing an API change."}
    ],
    max_tokens=256,
)
print(response.choices[0].message.content)

javascript

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.AVALAI_API_KEY,
  baseURL: "https://api.avalai.ir/v1",
  timeout: 60000,
  maxRetries: 0,
});
const response = await client.chat.completions.create({
  model: "muse-spark-open",
  messages: [{ role: "user", content: "Give a concise checklist for reviewing an API change." }],
  max_tokens: 256,
});
console.log(response.choices[0].message.content);

Illustrative response:

json

{
  "id": "chatcmpl-example-spark-open",
  "object": "chat.completion",
  "created": 1792886400,
  "model": "muse-spark-open",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Check compatibility, authentication, validation, errors, tests, and billing."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 20,
    "completion_tokens": 100,
    "total_tokens": 120,
    "prompt_tokens_details": {
      "cached_tokens": 0
    },
    "completion_tokens_details": null
  },
  "estimated_cost": {
    "unit": "0.0004500000",
    "irt": null,
    "exchange_rate": null
  }
}

This complete response is illustrative, not a live AvalAI receipt. Its ID, timestamp, answer, and token counts are synthetic. With 20 uncached input tokens and 100 output tokens, (20 × $1.25 + 100 × $4.25) / 1,000,000 = $0.0004500. The estimate applies no processing fee. irt and exchange_rate are null because no conversion rate was supplied. Actual reasoning/output accounting and optional usage or estimated_cost fields may differ or be absent; the User API remains the authoritative billing source.

Read choices[0].message.content and finish_reason. If the finish reason is length or the answer is empty, reassess the generation budget within independently checked model limits; the 256-token cap is deliberately small. For HTTP 400, fix the model ID, message structure, unsupported parameter, or oversized input. For 401/403, correct the key or account access. For 429, inspect the body and rate-limit guidance; exhausted account quota needs intervention. The published programs do not retry. Keep a sanitized error body and avalai-request-id when available, and remember a timeout can follow server-side processing before repeating a request.

See the October 25 announcement for the other models in this addition.

muse-glimmer-30b

Developed by Meta Superintelligence Lab and hosted by Fireworks.ai, Muse Glimmer is an approximately 30B-parameter dense causal model with a perception encoder. It combines text and image understanding, multilingual generation, tool use, multi-step reasoning, failure recovery, and a context window of at least 131,072 tokens.

Features

Text and image input with text output
More than 100 supported languages
Agentic task completion, coding, and tool use
Configurable reasoning strengths: low, medium, high, and xhigh
Recommended upstream sampling: temperature: 1.0, top_p: 0.95, top_k: 64

Pricing

Type	Cost per 1M tokens
Input	$0.35
Cached input	$0.04
Output	$1.50

Endpoint Support

Endpoint	Support
`v1/chat/completions`	✅ Full
`v1/messages`	✅ Full
`v1/responses`	⚠️ Partial

Example

BashPythonJavaScript

bash

curl https://api.avalai.ir/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -d '{
    "model": "muse-glimmer-30b",
    "messages": [{"role": "user", "content": "Plan a reliable tool-using research workflow."}]
  }'

python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AVALAI_API_KEY"], base_url="https://api.avalai.ir/v1"
)
response = client.chat.completions.create(
    model="muse-glimmer-30b",
    messages=[
        {"role": "user", "content": "Plan a reliable tool-using research workflow."}
    ],
)
print(response.choices[0].message.content)

javascript

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.AVALAI_API_KEY, baseURL: "https://api.avalai.ir/v1" });
const response = await client.chat.completions.create({
  model: "muse-glimmer-30b",
  messages: [{ role: "user", content: "Plan a reliable tool-using research workflow." }],
});
console.log(response.choices[0].message.content);

nemotron-3.5-lightning

Developed by NVIDIA and hosted by Fireworks.ai, Nemotron 3.5 Lightning is a 30B-total, 3B-active hybrid Mixture-of-Experts model. It interleaves Mamba-2, MoE, and attention layers for efficient agentic reasoning, coding, RAG, structured output, and tool use. AvalAI records a 262,144-token context window for this route.

Features

Efficient 30B-total / 3B-active hybrid MoE architecture
Reasoning, coding, RAG, and autonomous-agent workloads
Tool calling and structured output
Upstream thinking can be enabled or disabled
Recommended upstream sampling: temperature: 1.0, top_p: 0.95

Pricing

Type	Cost per 1M tokens
Input	$0.05
Cached input	$0.01
Output	$0.20

Endpoint Support

Endpoint	Support
`v1/chat/completions`	✅ Full
`v1/messages`	✅ Full
`v1/responses`	⚠️ Partial

Example

BashPythonJavaScript

bash

curl https://api.avalai.ir/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -d '{
    "model": "nemotron-3.5-lightning",
    "messages": [{"role": "user", "content": "Review this deployment plan and list the three highest risks."}]
  }'

python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AVALAI_API_KEY"], base_url="https://api.avalai.ir/v1"
)
response = client.chat.completions.create(
    model="nemotron-3.5-lightning",
    messages=[
        {
            "role": "user",
            "content": "Review this deployment plan and list the three highest risks.",
        }
    ],
)
print(response.choices[0].message.content)

javascript

import OpenAI from "openai";

const client = new OpenAI({ apiKey: process.env.AVALAI_API_KEY, baseURL: "https://api.avalai.ir/v1" });
const response = await client.chat.completions.create({
  model: "nemotron-3.5-lightning",
  messages: [{ role: "user", content: "Review this deployment plan and list the three highest risks." }],
});
console.log(response.choices[0].message.content);

nemotron-3-ultra

NVIDIA's large-scale Nemotron model is hosted on Fireworks.ai for complex reasoning, high-quality generation, and multi-step tool workflows.

Pricing

Type	Cost per 1M tokens
Input	$0.60
Cached input	$0.12
Output	$2.40

Endpoint Support

Endpoint	Support
`v1/chat/completions`	✅ Full
`v1/messages`	✅ Full
`v1/responses`	⚠️ Partial

Fireworks.ai ​

Available Models ​

API Endpoint Support ​

Key Features ​

Muse Spark Open ​

Access, Limits, and Choosing the Route ​

AvalAI Pricing and Prompt Caching ​

Bounded Chat Completions Example ​

muse-glimmer-30b ​

Features ​

Pricing ​

Endpoint Support ​

Example ​

nemotron-3.5-lightning ​

Features ​

Pricing ​

Endpoint Support ​

Example ​

nemotron-3-ultra ​

Pricing ​

Endpoint Support ​

Related Resources ​

Fireworks.ai

Available Models

API Endpoint Support

Key Features

Muse Spark Open

Access, Limits, and Choosing the Route

AvalAI Pricing and Prompt Caching

Bounded Chat Completions Example

muse-glimmer-30b

Features

Pricing

Endpoint Support

Example

nemotron-3.5-lightning

Features

Pricing

Endpoint Support

Example

nemotron-3-ultra

Pricing

Endpoint Support

Related Resources