feat: upgrade MiniMax default model to M3#3089
Conversation
Add MiniMax (MiniMax-M2.5 and MiniMax-M2.5-highspeed) as a new LLM provider option alongside OpenRouter. MiniMax offers 204K context window models via an OpenAI-compatible API. Changes: - Add MINIMAX enum and models to LLMProvider definitions - Add MiniMax provider initialization using @ai-sdk/openai-compatible - Add @ai-sdk/openai-compatible dependency for OpenAI-compatible providers - Add MINIMAX_API_KEY as optional env var - Update .env.example and self-hosting docs
|
Someone is attempting to deploy a commit to the Onlook Team on Vercel. A member of the Team first needs to authorize it. |
📝 WalkthroughWalkthroughAdds MiniMax as an optional LLM provider: environment variable exposure, new Minimax model enums and token limits, OpenAI-compatible provider integration (dependency added), and documentation updates listing MiniMax as a built‑in provider. ChangesMiniMax provider
Sequence Diagram(s)sequenceDiagram
participant Client as Client
participant Server as Server (app)
participant Provider as Minimax Provider Adapter
participant API as MiniMax API
Client->>Server: request LLM completion
Server->>Provider: select model & attach MINIMAX_API_KEY
Provider->>API: HTTP request to https://api.minimax.io/v1 (OpenAI-compatible)
API-->>Provider: completion response
Provider-->>Server: normalized response
Server-->>Client: deliver completion
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
packages/ai/src/chat/providers.ts (1)
38-40: Consider adding headers or providerOptions for tracking.The OpenRouter case sets
headerswithHTTP-RefererandX-Titlefor tracking/attribution purposes. If MiniMax supports similar headers, consider adding them for consistency and attribution.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/ai/src/chat/providers.ts` around lines 38 - 40, The MiniMax branch (case LLMProvider.MINIMAX) currently only calls getMinimaxProvider(requestedModel) without attaching tracking headers/providerOptions; mirror the OpenRouter handling by passing through providerOptions or headers (e.g., HTTP-Referer and X-Title) when constructing the MiniMax provider so attribution/tracking is included—update the getMinimaxProvider invocation or its returned config to accept and forward a headers/providerOptions object (match the shape used in the OpenRouter case) and ensure LLMProvider.MINIMAX uses that headers/providerOptions.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@packages/ai/package.json`:
- Line 36: The package.json entry for the dependency "@ai-sdk/openai-compatible"
uses a nonexistent version ("^1.0.34"); update that dependency in
packages/ai/package.json by replacing the version string with a valid released
version such as "^2.0.9" (or "^2.0.8" if preferred) so the dependency resolves
correctly during installs.
---
Nitpick comments:
In `@packages/ai/src/chat/providers.ts`:
- Around line 38-40: The MiniMax branch (case LLMProvider.MINIMAX) currently
only calls getMinimaxProvider(requestedModel) without attaching tracking
headers/providerOptions; mirror the OpenRouter handling by passing through
providerOptions or headers (e.g., HTTP-Referer and X-Title) when constructing
the MiniMax provider so attribution/tracking is included—update the
getMinimaxProvider invocation or its returned config to accept and forward a
headers/providerOptions object (match the shape used in the OpenRouter case) and
ensure LLMProvider.MINIMAX uses that headers/providerOptions.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 6e2a1e7e-48bb-4a24-b87a-662df11cdaa3
📒 Files selected for processing (6)
apps/web/client/.env.exampleapps/web/client/src/env.tsdocs/content/docs/self-hosting/external-services.mdxpackages/ai/package.jsonpackages/ai/src/chat/providers.tspackages/models/src/llm/index.ts
| "typescript": "^5.5.4" | ||
| }, | ||
| "dependencies": { | ||
| "@ai-sdk/openai-compatible": "^1.0.34", |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Check available versions of `@ai-sdk/openai-compatible`
curl -s https://registry.npmjs.org/@ai-sdk/openai-compatible | jq '.versions | keys | .[-5:]'
# Check peer dependencies
curl -s https://registry.npmjs.org/@ai-sdk/openai-compatible/1.0.34 | jq '.peerDependencies'Repository: onlook-dev/onlook
Length of output: 173
Update @ai-sdk/openai-compatible to a valid version.
Version 1.0.34 does not exist in the npm registry. The latest stable versions are 2.0.8 and 2.0.9. Update the dependency to one of these versions or the latest available version that meets the project requirements.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/ai/package.json` at line 36, The package.json entry for the
dependency "@ai-sdk/openai-compatible" uses a nonexistent version ("^1.0.34");
update that dependency in packages/ai/package.json by replacing the version
string with a valid released version such as "^2.0.9" (or "^2.0.8" if preferred)
so the dependency resolves correctly during installs.
- Add MiniMax-M2.7 and MiniMax-M2.7-highspeed to model list - Set MiniMax-M2.7 as default model (first in enum) - Keep all previous models (M2.5, M2.5-highspeed) as alternatives - Update docs to reference M2.7
octo-patch
changed the title
There was a problem hiding this comment.
🧹 Nitpick comments (1)
packages/models/src/llm/index.ts (1)
43-53: Consider compile-time exhaustiveness forMODEL_MAX_TOKENS.Future model additions can miss a token entry silently. Typing the map as a
Record<OPENROUTER_MODELS | MINIMAX_MODELS, number>will make omissions a type error.♻️ Proposed refactor
-export const MODEL_MAX_TOKENS = { +export const MODEL_MAX_TOKENS: Record<OPENROUTER_MODELS | MINIMAX_MODELS, number> = { [OPENROUTER_MODELS.CLAUDE_4_5_SONNET]: 200000, [OPENROUTER_MODELS.CLAUDE_3_5_HAIKU]: 200000, [OPENROUTER_MODELS.OPEN_AI_GPT_5_NANO]: 400000, [OPENROUTER_MODELS.OPEN_AI_GPT_5_MINI]: 400000, [OPENROUTER_MODELS.OPEN_AI_GPT_5]: 400000, [MINIMAX_MODELS.MINIMAX_M2_7]: 204000, [MINIMAX_MODELS.MINIMAX_M2_7_HIGHSPEED]: 204000, [MINIMAX_MODELS.MINIMAX_M2_5]: 204000, [MINIMAX_MODELS.MINIMAX_M2_5_HIGHSPEED]: 204000, -} as const; +};🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/models/src/llm/index.ts` around lines 43 - 53, MODEL_MAX_TOKENS is typed too loosely so adding new models can silently omit entries; change its declaration to have an explicit compile-time exhaustive type such as Record<OPENROUTER_MODELS | MINIMAX_MODELS, number> (replace the current inferred/`as const` typing) so the compiler will error when any member of OPENROUTER_MODELS or MINIMAX_MODELS is missing; update the constant name MODEL_MAX_TOKENS and ensure you provide numeric entries for every enum member from OPENROUTER_MODELS and MINIMAX_MODELS to satisfy the new type.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In `@packages/models/src/llm/index.ts`:
- Around line 43-53: MODEL_MAX_TOKENS is typed too loosely so adding new models
can silently omit entries; change its declaration to have an explicit
compile-time exhaustive type such as Record<OPENROUTER_MODELS | MINIMAX_MODELS,
number> (replace the current inferred/`as const` typing) so the compiler will
error when any member of OPENROUTER_MODELS or MINIMAX_MODELS is missing; update
the constant name MODEL_MAX_TOKENS and ensure you provide numeric entries for
every enum member from OPENROUTER_MODELS and MINIMAX_MODELS to satisfy the new
type.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 273d81ee-2d40-46c5-b417-243b31cadb19
📒 Files selected for processing (2)
docs/content/docs/self-hosting/external-services.mdxpackages/models/src/llm/index.ts
- Add MiniMax-M3 to model list and set as default - Keep MiniMax-M2.7 and MiniMax-M2.7-highspeed - Remove older models (M2.5/M2.1/M2/M1) - Update related tests and docs Co-Authored-By: Octopus <liyuan851277048@icloud.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@packages/models/src/llm/index.ts`:
- Around line 48-50: The MINI MAX model token limits in the MODEL_MAX_TOKENS
mapping are set below the official context-window limits for MINIMAX_M3 and
MINIMAX_M2_7/M2_7_HIGHSPEED; update the entries keyed by
MINIMAX_MODELS.MINIMAX_M3, MINIMAX_MODELS.MINIMAX_M2_7, and
MINIMAX_MODELS.MINIMAX_M2_7_HIGHSPEED to the official maxima (set M3 to 1000000
with note that 512000 is guaranteed minimum, and set M2.7 / M2.7-highspeed to
204800) so that consumers using MODEL_MAX_TOKENS (e.g., maxOutputTokens in
packages/ai/src/chat/providers.ts) get the correct budget, or alternatively
rename/document the mapping to indicate these are "guaranteed minimum" values
rather than maxima if you intend to keep the lower numbers.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 5c42584e-995e-4245-a403-246e5cc5c3da
📒 Files selected for processing (2)
docs/content/docs/self-hosting/external-services.mdxpackages/models/src/llm/index.ts
✅ Files skipped from review due to trivial changes (1)
- docs/content/docs/self-hosting/external-services.mdx
| [MINIMAX_MODELS.MINIMAX_M3]: 512000, | ||
| [MINIMAX_MODELS.MINIMAX_M2_7]: 204000, | ||
| [MINIMAX_MODELS.MINIMAX_M2_7_HIGHSPEED]: 204000, |
There was a problem hiding this comment.
🧩 Analysis chain
🌐 Web query:
What are the maximum context window sizes (token limits) for MiniMax M3, M2.7, and M2.7-highspeed models?
💡 Result:
The maximum context window sizes (token limits) for the specified MiniMax models are as follows: MiniMax M3: Supports a context window of 1,000,000 tokens [1][2][3]. MiniMax M2.7: Supports a context window of 204,800 tokens [2][4][5]. MiniMax M2.7-highspeed: Supports a context window of 204,800 tokens [2][6][7]. These limits reflect the official specifications provided in MiniMax's documentation and related technical resources [2][8][9].
Citations:
- 1: https://www.minimax.io/models/text/m3
- 2: https://platform.minimax.io/docs/guides/text-generation
- 3: https://aimlapi.com/models/minimax-m3
- 4: https://docs.api.nvidia.com/nim/reference/minimaxai-minimax-m2.7
- 5: https://www.llmreference.com/model/minimax-m2.7
- 6: https://vercel.com/ai-gateway/models/minimax-m2.7-highspeed
- 7: https://www.llmreference.com/model/minimax-m2.7-highspeed
- 8: https://www.minimax.io/blog/minimax-m3
- 9: https://developer.nvidia.com/blog/minimax-m2-7-advances-scalable-agentic-workflows-on-nvidia-platforms-for-complex-ai-applications/
🌐 Web query:
What does the official MiniMax documentation state for the maximum context window size (token limit) of the MiniMax text generation model M3?
💡 Result:
According to the official MiniMax documentation, the MiniMax M3 text generation model supports a context window of 1,000,000 tokens [1][2]. The model leverages a proprietary architecture called MiniMax Sparse Attention (MSA) to achieve this capacity [1][3]. Some documentation also specifies that the model maintains a guaranteed minimum of 512,000 tokens within this context window [1][4]. Additionally, the model is noted to support a maximum output of 1,000,000 tokens per request [5][6].
Citations:
- 1: https://www.minimax.io/models/text/m3
- 2: https://platform.minimax.io/docs/guides/text-generation
- 3: https://www.minimax.io/blog/minimax-m3
- 4: https://lushbinary.com/blog/minimax-m3-developer-guide-benchmarks-pricing-msa-architecture/
- 5: https://vercel.com/ai-gateway/models/minimax-m3/faq
- 6: https://vercel.com/ai-gateway/models/minimax-m3/about
🌐 Web query:
What does the official MiniMax documentation state for the maximum context window size (token limit) of the MiniMax text generation model M2.7?
💡 Result:
According to the official MiniMax API documentation, the MiniMax M2.7 text generation model has a maximum context window size of 204,800 tokens [1][2][3]. This limit applies to the total number of input and output tokens combined [1]. Authoritative sources, including the MiniMax platform documentation and technical integrations, consistently cite this 204,800-token (often referred to as 200K) capacity for the M2.7 series [4][5][1][2]. While some local deployment or alternative environment documentation may reference specific usage constraints or rounded figures [6][7], the official API specifications maintain 204,800 as the standard context limit [1][3].
Citations:
- 1: https://platform.minimax.io/docs/api-reference/api-overview
- 2: https://platform.minimax.io/docs/api-reference/text-openai-api
- 3: https://platform.minimax.io/docs/guides/text-generation
- 4: https://docs.api.nvidia.com/nim/reference/minimaxai-minimax-m2.7
- 5: https://developer.nvidia.com/blog/minimax-m2-7-advances-scalable-agentic-workflows-on-nvidia-platforms-for-complex-ai-applications/
- 6: https://platform.minimax.io/docs/guides/local-deploy
- 7: https://unsloth.ai/docs/models/tutorials/minimax-m27
🌐 Web query:
What does the official MiniMax documentation state for the maximum context window size (token limit) of the MiniMax text generation model M2.7-highspeed?
💡 Result:
According to the official MiniMax API documentation, the MiniMax text generation model M2.7-highspeed has a context window size of 204,800 tokens [1][2][3]. This maximum token count refers to the combined total of both input and output tokens [2].
Citations:
- 1: https://platform.minimax.io/docs/guides/text-generation
- 2: https://platform.minimax.io/docs/api-reference/api-overview
- 3: https://platform.minimax.io/docs/api-reference/text-ai-sdk
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Show the relevant portion of the file around the snippet
sed -n '1,140p' packages/models/src/llm/index.ts | nl -ba | sed -n '35,90p'
# Find the constant identifier(s) around those lines
rg -n "MINIMAX_MODELS|MAX_TOKENS|token" packages/models/src/llm/index.ts
# Show where MAX_TOKENS values are used (across repo) to interpret semantics
# (Limit search to minimax model usage)
rg -n "MINIMAX_MODELS.*M3|MINIMAX_M2_7|MINIMAX_M2_7_HIGHSPEED" -S packages/models/src/llmRepository: onlook-dev/onlook
Length of output: 104
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Show relevant section with line numbers (nl not available)
cat -n packages/models/src/llm/index.ts | sed -n '35,90p'
# Locate related constants / token limit definitions
rg -n "MINIMAX_MODELS|MAX_TOKENS|token" packages/models/src/llm/index.ts
# Find where the specific MINIMAX_MODELS token values are used
rg -n "MINIMAX_MODELS\.MINIMAX_M3|MINIMAX_MODELS\.MINIMAX_M2_7|MINIMAX_MODELS\.MINIMAX_M2_7_HIGHSPEED" -S packages/models/src/llm
# Also search across repo in case values are referenced elsewhere
rg -n "MINIMAX_MODELS\.MINIMAX_M3|MINIMAX_MODELS\.MINIMAX_M2_7|MINIMAX_MODELS\.MINIMAX_M2_7_HIGHSPEED" -S .Repository: onlook-dev/onlook
Length of output: 1540
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Find usages of MODEL_MAX_TOKENS
rg -n "MODEL_MAX_TOKENS" -S .
# Find how maxOutputTokens / token limits are used around providers
rg -n "maxOutputTokens" -S packages/models/src/llm packages -S
# Specifically see where token limits are enforced (truncate/chunk)
rg -n "truncate|truncat|token( |-|_)limit|max.*token|max_tokens|maxTokens|context window" -S packages/models/src/llm packages/models/srcRepository: onlook-dev/onlook
Length of output: 949
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "---- packages/ai/src/chat/providers.ts ----"
cat -n packages/ai/src/chat/providers.ts | sed -n '1,140p'
echo
echo "---- chat-context.tsx ----"
cat -n apps/web/client/src/app/project/[id]/_components/right-panel/chat-tab/chat-input/chat-context.tsx | sed -n '1,140p'
# Look for minimax request payload fields
echo
echo "---- Search for minimax request max_tokens fields ----"
rg -n "minimax|MINIMAX|max_tokens|maxTokens|maxOutputTokens|context|tokens" -S packages/ai/src/chat packages/ai/srcRepository: onlook-dev/onlook
Length of output: 18637
🏁 Script executed:
#!/bin/bash
set -euo pipefail
rg -n "maxOutputTokens" -S packages/ai/src
# Show likely call sites: files that call initModel or consume ModelConfig
rg -n "initModel\\(" -S packages/ai/src
# Follow MODEL_MAX_TOKENS usage further
rg -n "MODEL_MAX_TOKENS\\[" -S packages/ai/src packages -SRepository: onlook-dev/onlook
Length of output: 701
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "---- packages/ai/src/agents/root.ts ----"
cat -n packages/ai/src/agents/root.ts | sed -n '1,180p'
echo
echo "---- usages of maxOutputTokens in packages/ai/src ----"
rg -n "maxOutputTokens" -S packages/ai/src
echo
echo "---- search for maxTokens / max_tokens in packages/ai/src ----"
rg -n "maxTokens|max_tokens|max_output_tokens" -S packages/ai/srcRepository: onlook-dev/onlook
Length of output: 5197
🏁 Script executed:
#!/bin/bash
set -euo pipefail
rg -n "maxOutputTokens" -S .
rg -n "maxTokens\\b|max_tokens\\b|maxOutput\\b" -S packages appsRepository: onlook-dev/onlook
Length of output: 1960
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "---- chat/conversation.ts ----"
cat -n apps/web/client/src/server/api/routers/chat/conversation.ts | sed -n '1,200p'
echo
echo "---- chat/suggestion.ts ----"
cat -n apps/web/client/src/server/api/routers/chat/suggestion.ts | sed -n '1,120p'
echo
echo "---- project.ts around line ~329 ----"
cat -n apps/web/client/src/server/api/routers/project/project.ts | sed -n '280,380p'
echo
echo "---- search packages/ai for maxOutputTokens usage ----"
rg -n "maxOutputTokens" -S packages/ai/srcRepository: onlook-dev/onlook
Length of output: 13385
🏁 Script executed:
#!/bin/bash
set -euo pipefail
rg -n "LLMProvider\\.MINIMAX|MINIMAX_MODELS\\." -S .Repository: onlook-dev/onlook
Length of output: 466
Align MiniMax MODEL_MAX_TOKENS values with the official context-window limits.
- MiniMax M3: official max context window is 1,000,000 tokens (with a guaranteed minimum of 512,000), but code sets 512,000.
- MiniMax M2.7 / M2.7-highspeed: official max context window is 204,800 tokens, but code sets 204,000.
MODEL_MAX_TOKENSis used asmaxOutputTokens(packages/ai/src/chat/providers.ts), so these underestimates will reduce the usable budget when MiniMax is actually selected; either update to the official maxima or rename/document these constants as “safe/guaranteed minimum” rather than max.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@packages/models/src/llm/index.ts` around lines 48 - 50, The MINI MAX model
token limits in the MODEL_MAX_TOKENS mapping are set below the official
context-window limits for MINIMAX_M3 and MINIMAX_M2_7/M2_7_HIGHSPEED; update the
entries keyed by MINIMAX_MODELS.MINIMAX_M3, MINIMAX_MODELS.MINIMAX_M2_7, and
MINIMAX_MODELS.MINIMAX_M2_7_HIGHSPEED to the official maxima (set M3 to 1000000
with note that 512000 is guaranteed minimum, and set M2.7 / M2.7-highspeed to
204800) so that consumers using MODEL_MAX_TOKENS (e.g., maxOutputTokens in
packages/ai/src/chat/providers.ts) get the correct budget, or alternatively
rename/document the mapping to indicate these are "guaranteed minimum" values
rather than maxima if you intend to keep the lower numbers.
1 participant
Summary
Upgrade MiniMax model configuration to use M3 as the default model.
Changes
Why
MiniMax-M3 is the latest model, with a 512K context window, up to 128K output, and image input support.
Testing