Kimi K3.1 Pro, MiniMax M3.1 Flash, and New Muse Models
Date: 2026-10-25 / (1405-08-03)
Summary
AvalAI now offers Kimi K3.1 Pro and Highspeed, MiniMax M3.1 Flash, Meta Muse Wonder, and Muse Spark Open. These five request IDs provide choices for coding, reasoning, and long-context work. Their access tiers, token limits, and billed rates are listed below, with a shared Chat Completions example.
Details
Moonshot AI: Kimi K3.1 Pro and Highspeed
kimi-k3.1-prois Moonshot AI's K3.1 Pro model. The developer describes a post-trained Kimi K3 model for coding, knowledge work, visual understanding, and tasks that require many tool steps. It is a candidate for reviewing a codebase, planning a migration, or analyzing documents; evaluate the resulting answers on your own workload.kimi-k3.1-pro-highspeedis the higher-priced Highspeed option. Moonshot AI describes the same intelligence with approximately 2–3× faster generation. This is the developer's comparison, not an AvalAI latency measurement or speed guarantee.
The developer's release material is dated October 24, 2026. It describes weights planned for release by November 4, 2026; this announcement does not establish that the weights are already downloadable. Both IDs require AvalAI tier 1 or above. See the Kimi Pro and Highspeed provider sections.
MiniMax: M3.1 Flash
minimax-m3.1-flash adds MiniMax's M3.1 Flash model, released by the developer on October 19, 2026. MiniMax positions it for coding, tool use, agents, and long sessions, with text, image, and video understanding described in its upstream documentation. Its published generation-speed measurements are developer results; throughput and latency through AvalAI depend on the request and service conditions.
MiniMax's model-specific documentation describes adaptive reasoning that is always enabled upstream. The launch says model weights will be released later, while another guide uses the term “open-weight”; released-weight availability remains unresolved in those sources. The AvalAI ID is available from tier 0. Use the MiniMax provider section for the model's integration notes.
Meta: Muse Wonder and Muse Spark Open
muse-wonder-1.0is Meta's preview model for complex engineering and agents that plan, delegate, and work through long tasks. It produces text, including code and analysis; it is not an image-generation model. The separate Muse Wonder Full option in Meta's Muse Code subscription is not the AvalAI ID announced here. Access starts at tier 1. See Muse Wonder on AvalAI.muse-spark-openis the Meta-developed open-weight Muse Spark Open model, hosted through Fireworks AI on this AvalAI route. Meta describes a mixture-of-experts model with 816 billion total parameters and 42 billion active parameters, suited to reasoning, coding, tool-driven work, and visual understanding. Meta is the model developer; the hosting route does not change that attribution. Access starts at tier 0. See Muse Spark Open through Fireworks AI.
The developer descriptions above explain model capabilities. They do not establish an AvalAI video or audio interface, computer-control service, tool executor, or every native provider API option. The common integration shown in this announcement is POST https://api.avalai.ir/v1/chat/completions, using text messages and the exact AvalAI model ID. Consult the linked provider and Chat Completions reference before adding optional fields or choosing another endpoint.
Access and token limits
These are the separately recorded AvalAI catalog limits, in tokens:
| AvalAI request ID | Minimum tier | Input ceiling | Output ceiling |
|---|---|---|---|
kimi-k3.1-pro | 1 | 262144 | 262144 |
kimi-k3.1-pro-highspeed | 1 | 262144 | 262144 |
minimax-m3.1-flash | 0 | 1048576 | 1048576 |
muse-wonder-1.0 | 1 | 1048576 | 131072 |
muse-spark-open | 0 | 1048576 | Not recorded |
Do not add input and output ceilings together to infer a guaranteed combined context capacity, or assume both maxima can be used in one request. The input, requested output budget, reasoning usage, and model's shared context constraints must fit together. An unrecorded output ceiling is unknown, not unlimited. Account-specific request and token rate limits also apply; see rate limits.
The upstream limits differ in two important cases:
- Kimi documents a 1,048,576-token shared context upstream. AvalAI currently records 262,144 for each Kimi input and output ceiling; use the AvalAI limits for these routes.
- MiniMax documents a 1,000,000-token context and recommends 131,072 generation tokens, with an upstream generation maximum of 524,288. Those figures differ from AvalAI's separately recorded 1,048,576 input and output ceilings. The catalog's output value is not a verified upstream generation guarantee; keep the discrepancy in mind when sizing a real request.
Pricing: AvalAI base rates and billed rates
All token prices below are in USD per 1,000,000 tokens. “Cached input” is the rate for input actually accounted for as a cache hit. “Cache creation” is a separate category where listed; it is not an extra charge on every input token.
AvalAI base rates before the model-specific processing fee:
| Model | Input | Cached input | Cache creation | Output |
|---|---|---|---|---|
kimi-k3.1-pro | $2.00 | $0.20 | $4.00 | $10.00 |
kimi-k3.1-pro-highspeed | $4.00 | $0.20 | $8.00 | $20.00 |
minimax-m3.1-flash | $0.15 | $0.015 | Not separately listed | $0.60 |
muse-wonder-1.0 | $1.25 | $0.15 | Not separately listed | $4.25 |
muse-spark-open | $1.25 | $0.15 | Not separately listed | $4.25 |
Effective AvalAI billed rates, including the processing fee:
| Model | Processing fee | Input | Cached input | Cache creation | Output |
|---|---|---|---|---|---|
kimi-k3.1-pro | 0% | $2.00 | $0.20 | $4.00 | $10.00 |
kimi-k3.1-pro-highspeed | 15% | $4.60 | $0.23 | $9.20 | $23.00 |
minimax-m3.1-flash | 0% | $0.15 | $0.015 | Not separately listed | $0.60 |
muse-wonder-1.0 | 15% | $1.4375 | $0.1725 | Not separately listed | $4.8875 |
muse-spark-open | 0% | $1.25 | $0.15 | Not separately listed | $4.25 |
The 15% processing fee applies only to kimi-k3.1-pro-highspeed and muse-wonder-1.0 among these five additions. Multiply their base rates by 1.15 once. The prices in AvalAI's catalog and pricing page already contain this fee; do not add it again when estimating a bill. No new processing fee is applied to Pro, Flash, or Spark Open in this announcement.
These are AvalAI's supplied base and billed rates. Meta's current direct Model API listing for Muse Wonder shows $4.00 input, $0.20 cached input, and $20.00 output per 1 million tokens, which differs from the AvalAI base quote above. Do not relabel the AvalAI quote as Meta's current official price or derive AvalAI billing from that separate service's table.
How caching affects a bill
Use the token categories returned for the actual request. Cached reads, newly created cache entries, and ordinary uncached input represent different accounting categories; do not charge the same token as both ordinary input and a cache read or write. Output, including billed reasoning tokens where applicable, is charged separately. The ordinary text examples below have no cache creation or cache hits in their illustrative accounting.
Kimi's upstream one-hour cache-write prices match the base cache-creation figures above. That numeric match does not promise a one-hour AvalAI cache lifetime or support for a particular cache-control field. MiniMax documents automatic prefix caching upstream, for requests with at least 512 input tokens, with new input at the ordinary input rate and hits at the cached-input rate. This does not guarantee an AvalAI cache hit or lifetime. For Flash, Wonder, and Spark Open, no separate cache-creation rate is listed in this announcement; absence of a line does not mean that every request is cached or that creation is free.
API request and response examples
One text request, five exact model IDs
Set AVALAI_API_KEY in your server-side environment. The SDK tabs use the OpenAI Python or OpenAI JavaScript package, configured with AvalAI's /v1 base URL. The curl request uses a 10-second connection timeout, a 60-second total timeout, and no automatic retry. Both SDK clients use a 60-second timeout and disable their internal retries. Each program makes one request.
This small example uses max_tokens: 256 as the Chat Completions output budget. Reasoning models may spend that budget on reasoning and return little or no final text, or finish with length. Use a larger budget and an appropriate deadline for real tasks, within the chosen route's limits. The final answer and stop result shown later are illustrative, not guaranteed by this prompt.
curl --silent --show-error --fail-with-body \
--connect-timeout 10 --max-time 60 --retry 0 \
https://api.avalai.ir/v1/chat/completions \
-H "Authorization: Bearer $AVALAI_API_KEY" \
-H 'Content-Type: application/json' \
--data-binary '{
"model": "kimi-k3.1-pro",
"messages": [
{
"role": "user",
"content": "Give a concise checklist for reviewing an API change."
}
],
"max_tokens": 256
}'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,
)
completion = client.chat.completions.create(
model="kimi-k3.1-pro",
messages=[
{
"role": "user",
"content": "Give a concise checklist for reviewing an API change.",
}
],
max_tokens=256,
)
print(completion.model_dump_json(indent=2))import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.AVALAI_API_KEY,
baseURL: "https://api.avalai.ir/v1",
timeout: 60_000,
maxRetries: 0,
});
const completion = await client.chat.completions.create({
model: "kimi-k3.1-pro",
messages: [
{
role: "user",
content: "Give a concise checklist for reviewing an API change.",
},
],
max_tokens: 256,
});
console.log(JSON.stringify(completion, null, 2));The following table defines five distinct request payloads: use the complete example above and replace only its model string with the exact value in the chosen row. Keep the endpoint, message, and max_tokens: 256 unchanged. The same substitution applies in the curl, Python, and JavaScript tabs; the other four requests need no additional fields.
Exact model substitution | Illustrative USD cost: 20 uncached input + 100 output tokens |
|---|---|
kimi-k3.1-pro | $0.0010400000 |
kimi-k3.1-pro-highspeed | $0.0023920000 |
minimax-m3.1-flash | $0.0000630000 |
muse-wonder-1.0 | $0.0005175000 |
muse-spark-open | $0.0004500000 |
Complete illustrative response
This is a synthetic, illustrative non-streaming response, not a captured AvalAI response. Its 20 input and 100 output tokens are chosen to demonstrate billing; they were not measured with a tokenizer or a live request. With Pro's billed rates, (20 × 2.00 + 100 × 10.00) / 1000000 = $0.0010400000.
{
"id": "chatcmpl-illustrative-kimi-k31-pro",
"created": 1792886400,
"model": "kimi-k3.1-pro",
"object": "chat.completion",
"system_fingerprint": null,
"choices": [
{
"finish_reason": "stop",
"index": 0,
"message": {
"role": "assistant",
"content": "Check backward compatibility, request and response schemas, authentication, error handling, rate limits, tests, and documentation."
}
}
],
"usage": {
"prompt_tokens": 20,
"completion_tokens": 100,
"total_tokens": 120,
"prompt_tokens_details": {
"cached_tokens": 0
},
"completion_tokens_details": null
},
"estimated_cost": {
"unit": "0.0010400000",
"irt": null,
"exchange_rate": null
}
}Read final text from choices[0].message.content, and inspect finish_reason, usage, and estimated_cost. For the other four illustrative cases, substitute the corresponding model and estimated_cost.unit from the table; the chosen token counts stay the same only for this arithmetic example. Real responses can have different text, finish reasons, token counts, optional reasoning details, cache details, and cost fields. An echoed model ID does not independently identify an underlying version.
estimated_cost.unit shows the USD estimate in this illustration. irt and exchange_rate are null because no exchange-rate observation was made; unknown values must not be replaced with zero. For a real bill, use the returned accounting rather than these synthetic counts. The examples introduce basic text requests; they do not validate optional tools, structured output, vision inputs, streaming, or upstream reasoning controls through AvalAI.
Integration and troubleshooting
- Authentication and account access: keep the key on your server. For an authentication error, check the environment variable and the
Authorizationheader. For permission or model-access errors, verify the exact ID and the minimum tier in the table. - Invalid fields or oversized requests: inspect the response's error body and the chosen route's limits. Native Kimi and MiniMax documentation prefers
max_completion_tokensover its deprecatedmax_tokens; the published example uses AvalAI's common Chat Completions field. Do not copy native provider parameter rules or reasoning controls into this request without checking AvalAI compatibility. Upstream Kimi and Flash reasoning cannot be disabled by assuming anonesetting. - Empty or truncated answers: inspect
finish_reasonand any returned reasoning usage. Increase the output budget when reasoning consumed it, rather than treating empty final text as a valid answer. - 429 and temporary service failures: inspect the error code as well as the HTTP status. Rate limiting can justify a bounded retry after the indicated delay; terminal
quota_exceededorinsufficient_quotarequires account intervention. These examples deliberately make one attempt. - Timeouts and debugging: a timed-out request may already have been processed and charged. Preserve the safe error body, HTTP status, requested model, and
avalai-request-idresponse header when present before deciding whether to repeat it. Do not include API keys in logs or support reports. See error handling.
Evaluate each model on representative tasks, including Persian inputs where relevant. Validate returned code, calculations, and tool arguments in your application. Model output and provider benchmark claims do not replace permissions, deterministic checks, or human review for consequential actions.
Official sources
The provider documentation consulted for this announcement was retrieved on October 24, 2026. Provider capabilities and direct-provider prices are separate from the AvalAI access and billing facts above.
- Kimi: K3.1 Pro launch, quickstart, pricing, and context caching.
- MiniMax: M3.1 Flash launch, model invocation, Chat Completions schema, and prompt caching.
- Meta: model catalog, Muse Wonder, Muse Spark Open, and Spark Open deployment.