ISSUE DESCRIPTION Anthropic rejects duplicate websearch tool names When LibreChat web…
- User
ISSUE DESCRIPTION
Anthropic rejects duplicate web_search tool names
When LibreChat web search and Anthropic-native web search are both enabled, expose one
web_searchtool and prefer LibreChat's definition. Do not affect unrelated tools or other providers.Interface
No new public interfaces are introduced.
ADDITIONAL ADVICE
Since you are given a git repository, you can use git commands to simplify your work. For example, if you made a mistake and want to revert your changes, you can use
git checkout <file>to restore the file to its original state. You can also reset all changes in the repository usinggit reset --hardcommand. Additionally, you can usegit stashandgit stash popto temporarily save your changes and restore them later. Finally, you can see the changes that you've made (which will be submitted when you callsubmit) usinggit statusorgit diffcommands. However you don't need to usegit addorgit commitbefore submitting your changes. In fact,submitonly submits currently unstaged changes because it usesgit diffto get the changes that need to be submitted.CHECKLIST
Before submitting your solution, please go over the following checklist and make sure you've done everything:
- If an error was reported in the issue description, I have successfully reproduced it.
- If an error was reported in the issue description, I have confirmed that my fix resolves the error.
- I have thought about edge cases and made sure my fix handles them as well.
- I have run existing tests in the repository that might have been affected by the change I made and confirmed that they pass. I want you to list every bullet from this checklist and write your reasoning for why you think you did it or didn't need to.
Repository has been uploaded and your shell is currently at the repository root. Time to solve the issue!
(Current directory: /LibreChat, current file: none) bash-$
Assistant
I'll analyze the repository structure and understand the issue before making changes.
Assistant
Let me look at the key files for Anthropic-specific web search handling:
Assistant
Now let me look at the transform function or the place where tools are added to the Anthropic LLM:
Assistant
Now let me look at where the tools are merged/processed in the larger flow:
Assistant
Now let me look at the place where loading tools happens:
Assistant
Let me look at how tools are loaded:
Assistant
Let me look for where web_search is used in Anthropic flow:
Assistant
Let me look at the ToolService.js to understand the web_search handling:
Assistant
Now let me look at the Anthropic-specific LLM endpoint to see how the tools from LLM config get combined with the loaded tools:
Assistant
Let me check exactly how the tools are merged into the LLM config for the Anthropic endpoint:
Assistant
Let me look at the
@librechat/agentspackage since it handles the LLM creation:Assistant
Let me look at the Anthropic message preparation to understand how tools are merged:
Assistant
Let me look at the initialize.js for the agents endpoint:
Assistant
Let me look at the agent initialize function to see how tools end up in the LLM config:
Assistant
Now let's look at where the tools are finalized in the LLM config. Let me look at the Anthropic endpoint to find where the tools from LLM config (the
toolswe set ingetLLMConfig) get integrated:Assistant
Let me look at the test file for the Anthropic LLM to understand the expected behavior:
Assistant
Now I understand the test. Let me understand the issue completely. The problem is that when LibreChat's native web search tool is enabled (it's added to
agent.toolsas 'web_search'), and Anthropic's native web search is also enabled (viamodel_parameters.web_search), both create a tool withname: 'web_search', which Anthropic rejects.Let me find where the tools are merged:
Assistant
Now let me look at the handleTools.js to understand how the LibreChat web_search tool is constructed:
Assistant
Now let me check the Tools enum to understand its definition:
Assistant
Now let's understand the actual flow more clearly. Let me look at the test file for the
getLLMConfigto understand how the result is expected:Assistant
Now let me look at the test for "config.anthropic.spec.ts" which likely has the test for the endpoint:
Assistant
Now let me look at the config.anthropic.ts:
Assistant
Now let me look at the OpenAI llm.ts to see how OpenAI handles web_search:
Assistant
Now I have a clearer picture. Let me check the agent's
toolsarray - if both web_search (from agent.tools) and the anthropic-default web_search tool are active, they would produce duplicates.Let me look for how the tools are merged into the LLM config in the agent flow:
Assistant
Let me look at the createRun to find tool merging:
Assistant
Let me look at where the tools are added to the LLM invoke:
Assistant
Let me look at the @librechat/agents anthropic.ts to see how
web_searchis handled:Assistant
Let me look at @librechat/agents index.ts to understand the architecture:
Assistant
Let me look at the LibreChat
createSearchToolto understand what name it returns:Assistant
Now I have a clearer picture. Both tools are named
web_search:- LibreChat's
createSearchToolreturns a tool withname: 'web_search' - Anthropic's
getLLMConfigadds{ type: 'web_search_20250305', name: 'web_search' }when web_search is enabled
Let me look at the issue more carefully. Let me trace the flow:
- LibreChat's
Assistant
Now let me get the bigger picture by looking at the LibreChat agent's structure - how does the tool instantiate and end up conflicting?
Let me look at the
responses/types.tsto understand tool type:Assistant
Now let me look at the
added.tsto understand how tools are added:Assistant
Now let me dig into the actual issue. The Anthropic LLM has been configured with
web_search: trueso it adds a tool{ type: 'web_search_20250305', name: 'web_search' }. Then the agent loads the LibreChatweb_searchtool which is also namedweb_search.The fix should be in the agent's initialize.ts where it combines LibreChat's tools with the LLM-specific tools. Let me look at where the tools from the LLM config and the loaded tools are combined:
Assistant
Now I see the issue. Let me trace
options.toolsflow. Let me look at whereoptions.toolsis set in the agent initialize flow:Assistant
Now let me look at the entire initialize flow - how the
toolsfrom LLM config (the web_search tool) end up being passed to initializeAgent:Assistant
Now let me look at where the
toolsfrom the LLM config are added back to the agent's tools list. I need to see if the LLMweb_searchtool is added to the agent'stoolssomewhere:Assistant
Let me look at the agent's run.ts to see how tools are passed:
Assistant
Let me look at the
initializeAgentflow where the tools are bundled. Looking at the agent'stoolsarray initialized at line 829-846 of initialize.ts:Assistant
Now let me find where the
options.toolscome from. Theoptionshere isLoadAgentOptionsor similar. Let me check more:Assistant
Let me look at the structure of
optionsforinitializeAgent:Assistant
Now I see. The
options.toolscome fromgetOptionswhich callsgetProviderConfig. Let me look at whatgetOptionsreturns:Assistant
Let me look at the InitializeResultBase type:
Assistant
So
options.toolscomes from the Anthropic initialize which returnstools: [{ type: 'web_search_20250305', name: 'web_search' }].Now I need to understand the full flow. Let me look at how the agent's
toolsare set:Assistant
Now let me look at how the agent's
toolsarray (fromagent.tools) is used. The agent'stoolsarray contains 'web_search' (string), and theoptions.tools(from the LLM config) might contain{ type: 'web_search_20250305', name: 'web_search' }. These are later passed toloadToolswhich converts the string 'web_search' to a tool named 'web_search'.Let me trace how
options.tools(the LLM options) interacts withagent.tools(the string list):Assistant
Now let me look at what
toolDefinitions(structuredTools) is set to:Assistant
OK now I understand the flow. Let me look at this:
agent.toolsis a list of strings like['web_search', 'execute_code', ...]loadToolsis called with these strings and returnsstructuredTools(the LibreChat's web_search tool etc.) andtoolDefinitionsoptions.toolscomes from the LLM config (anthropic web_search tool fromgetLLMConfig)- The agent's
tools(line 829) is set tostructuredToolsifoptions.toolsis empty, otherwiseoptions.tools, otherwisestructuredTools.concat(options.tools)
So when both
options.tools(LLM-specific) andstructuredTools(LibreChat's tools) are present, they're concatenated. This is where the problem occurs.Let me look at the existing test to understand the expected behavior:
Assistant
Now let me look at the initialize.test.ts file to see how the agent initialization is tested:
Assistant
Now let me look at the test on line 745 which has
agent.tools = ['web_search']:Assistant
Now let me look at where the
toolsfrom LLM config are concatenated. Looking at the initialize.ts file:Assistant
Now I see the issue clearly. When:
agent.provider === Providers.ANTHROPICoptions.tools(LLM) has[{ type: 'web_search_20250305', name: 'web_search' }]structuredTools(LibreChat) has LibreChat'sweb_searchtool
Both will be concatenated, creating two
web_searchtools. This is the duplicate.Now let me think about the fix. The issue says:
When LibreChat web search and Anthropic-native web search are both enabled, expose one
web_searchtool and prefer LibreChat's definition. Do not affect unrelated tools or other providers.The fix should:
- Detect when both LibreChat's web_search AND Anthropic's web_search are enabled
- Only expose one of them (prefer LibreChat's)
- Not affect other providers or other tools
Let me see where both options.tools and structuredTools are added together. The duplicated logic is at the
tools = structuredTools.concat(options.tools)step.Let me look at the
initialize.test.tsto see how the tools are tested:Assistant
Now let me look at the createMocks function in tests:
Assistant
Now let me check the test file for the existing tests on the merge logic:
Assistant
Let me look at how the tools get into the LLM when bindTools is called. Looking at the
toolsfield initialization from the agent flow:Assistant
Now let me look at the agent's
toolsafter initialize. It is passed to theRun.createasagent.tools. Let me look at thecreateRunto check howtoolsis used:Assistant
Let me look at
this.toolsin AgentContext:Assistant
So
this.toolsis set from theagent.toolsfrom the initialize flow. Looking at the initialize.ts, theagent.toolsis concatenated withoptions.tools(LLM-specific tools). So when both LibreChat'sweb_searchAND Anthropic'sweb_searchare present, both are inagent.tools, and they both get bound to the LLM, causing the duplicate.Now let me look at the actual problem. Looking at the merge logic:
Assistant
So the issue is here. When:
agent.provider === Providers.ANTHROPICoptions.toolshas{ type: 'web_search_20250305', name: 'web_search' }structuredToolshas LibreChat's web_search tool (also namedweb_search)
Both tools get concatenated.
Now, the fix should be to deduplicate the tools by name, preferring LibreChat's tools over the LLM-provided tools. Let me look at the actual tool structure to understand:
Assistant
Now let me look at the existing tests for the concat logic - what's the test verifying?
Assistant
Now let me write a simple test to reproduce the issue first. Let me see how the test for the concat logic works:
Assistant
Now let me write a script to reproduce the issue:
Assistant
The issue is reproduced. Now let me think about the fix.
The fix should be in
initialize.tswhere the tools are concatenated. The condition should:- When both
options.tools(LLM-specific tools) andstructuredTools(LibreChat's tools) have the samename, prefer LibreChat's tool - Only affect the
ANTHROPICprovider (since that's the issue) - Not affect other providers or other tools
Let me also look at the
toolDefinitions(event-driven mode), since that's another path where tools are exposed:- When both
Assistant
Now let me look at the types more carefully - LCTool is the tool definition:
Assistant
Now I need to understand the structure of the tool. The
options.toolsitems are tool definitions (like{ type: 'web_search_20250305', name: 'web_search' }). ThestructuredToolsareGenericToolinstances (withnameproperty).Both have a
nameproperty. So I can deduplicate by name.Let me look at the structuredTools to confirm this:
Assistant
Good. Now let's look at the
StructuredToolInterfaceto understand the structure:Assistant
Good. Both
StructuredTool(LangChain) andBindToolsInputhave anameproperty. So both can be deduplicated byname.Now let me look at the issue more carefully. The fix should be in
initialize.tsand should:- Detect when both LibreChat's web_search AND Anthropic's web_search are enabled
- Only expose one of them (prefer LibreChat's)
- Not affect other providers or other tools
Let me design the fix:
Assistant
Now let me think about the fix. Now let me look at the
toolsproperty type more carefully:Assistant
Now let me look at the
toolstypes ininitialize.ts. Theoptions.toolsisunknown[]based on the type definition:Assistant
Now let me also check the
Toolsenum:Assistant
Now I need to understand the structure of the
options.toolsbetter. Theoptions.toolsfor the Anthropic endpoint returns:{ type: 'web_search_20250305', name: 'web_search' }And the
structuredToolsfor the LibreChat web search tool returns aDynamicStructuredToolwithname: 'web_search'.Let me write the fix. The cleanest approach is to deduplicate by name in the concat:
Assistant
Now let me think about the best place for the fix. Looking at the code:
let tools: GenericTool[] = options.tools?.length ? (options.tools as GenericTool[]) : (structuredTools ?? []); if ( (agent.provider === Providers.OPENAI || agent.provider === Providers.AZURE || agent.provider === Providers.ANTHROPIC) && options.tools?.length && structuredTools?.length ) { tools = structuredTools.concat(options.tools as GenericTool[]); }The fix should:
- When concatenating, dedupe by name
- Prefer LibreChat's tools (in
structuredTools)
Let me write the fix:
Assistant
Now let me also think about the toolDefinitions path - if the tool is in event-driven mode, the
toolDefinitionsare used:Assistant
Let me see how
toolDefinitionsandtoolsare related:Assistant
Now I see. In event-driven mode, both
toolDefinitions(fromloadTools) andtools(fromoptions.tools) are used. Thetoolsfrom options also need to be deduplicated. Let me look at the bindTools logic:Assistant
Let me look at the relevant code path more carefully:
Assistant
I see—so in event-driven mode, both
toolDefinitions(LibreChat's tools) andinstanceTools(fromthis.tools, which is the concatenated tools array) get bound. So if both the LibreChat's web_search and the Anthropic's web_search are in thetoolsarray, both will get bound.So the fix needs to:
- In the
toolsarray (the LLM bind tools), deduplicate by name when concatenating - The
toolsarray is what's passed tobindTools()
Let me also check that the fix doesn't affect other providers. The fix should be limited to the case where both LibreChat's tools and LLM-specific tools share the same name.
Let me also think about this more carefully. Looking at the tools flow:
agent.toolsis a String[] (e.g.,['web_search'])loadToolsconverts these strings intostructuredTools(which are tool instances likeDynamicStructuredTool)loadToolsalso returnstoolDefinitions(the LCTool schema for event-driven mode)options.toolsfromgetLLMConfigreturns provider-specific tools like{ type: 'web_search_20250305', name: 'web_search' }
The
options.toolsis the provider's native tool definitions (e.g., for Anthropic, theweb_search_20250305server tool). These are bound to the LLM directly.The
structuredToolsare the LibreChat-side tool instances (for direct execution).Both end up bound to the LLM (because the LLM needs to know about all available tools in its function-calling API). That's why duplicating the name causes the issue.
Now let me write the fix. Let me first look at the relevant code one more time:
- In the
Assistant
Now let me write the fix:
Assistant
Now let me think about where to put the fix. The simplest fix is to dedupe by name in the
toolsarray. Let me look at the issue once more:The issue is that when both LibreChat's
web_search(string) and Anthropic'sweb_search(LLM config) are enabled, both get added as tools withname: 'web_search'. The fix should be to only expose one. Per the issue, prefer LibreChat's.Let me write the fix. Looking at the code:
Assistant
Now I'll write the fix. The dedup should be done by name, preferring LibreChat's tools. Let me also check what
options.toolslooks like for both OpenAI and Anthropic:Assistant
So OpenAI's web_search tool is
{ type: 'web_search' }(no name). The Anthropic one is{ type: 'web_search_20250305', name: 'web_search' }.For OpenAI, the tool has
type: 'web_search'but noname. So the dedup needs to be careful here.Let me think about this more carefully. The issue says:
When LibreChat web search and Anthropic-native web search are both enabled, expose one
web_searchtool and prefer LibreChat's definition. Do not affect unrelated tools or other providers.So the fix should:
- Only apply when both LibreChat's web_search AND Anthropic's web_search are enabled
- Prefer LibreChat's definition
- Don't affect other tools or other providers
Let me look at the type for OpenAI web_search tool:
Assistant
The OpenAI version uses
{ type: 'web_search' }(no name). Hmm, let me check the OpenAI Responses API spec for this:Assistant
OK so the OpenAI web_search tool is recognized by its
typefield. The Anthropic one has bothtypeandname.Now let me focus on the fix. The issue is that we have two tools with the same name (LibreChat's
web_searchand Anthropic'sweb_search). For the Anthropic endpoint, we need to deduplicate.Looking at the existing code logic, the fix should be in the
else ifbranch where the tools are concatenated. The fix should:- Dedupe by name
- Prefer LibreChat's tools (from
structuredTools) - Only apply to the specific case where names match
Let me also consider that
options.toolsfor OpenAI is{ type: 'web_search' }(no name, so no conflict). The conflict only happens with Anthropic's{ type: 'web_search_20250305', name: 'web_search' }.Let me design the fix:
Assistant
Wait, for OpenAI, we'd still have a duplicate because the OpenAI tool is
{ type: 'web_search' }(no name). But this is a separate test case. Let me think about this carefully.Actually, looking at the OpenAI test, the final list contains both with different
name(calculator and web_search), and the LibreChat'sweb_searchis the LibreChat's tool, and the OpenAI'sweb_search(no name but type) is also there. So OpenAI doesn't have a name conflict on the surface, but they could potentially conflict when bound to the LLM.Let me check the OpenAI Responses API spec for both:
Assistant
OK so OpenAI's web search tool uses
type: 'web_search_preview' | 'web_search_preview_2025_03_11'- no name. So OpenAI doesn't have a name conflict.Now let me focus on the actual fix. The issue specifically says:
Anthropic rejects duplicate web_search tool names
So the fix should specifically deduplicate when both LibreChat's web_search and Anthropic's web_search are present. Let me look at the existing dedup logic - if any, from agent.tools:
Assistant
Let me look at the unionPrimeAllowedTools to see how it does dedup:
Assistant
Good. Now let me write the fix:
Assistant
All test cases pass. Now let me also think about the OpenAI scenario. Although the issue says it only affects Anthropic, the dedup logic should be safe for OpenAI too because tools without a
namefield won't be filtered out.Let me now write the actual fix:
Assistant
Now let me apply the fix:
Assistant
Let me verify the change was applied correctly:
Assistant
Now let me also verify that the existing tests still pass. Let me look at the existing tests for this specific code:
Assistant
These errors are about the modules not being built. Let me build the data-provider first:
Assistant
Now let me try the tsc again:
Assistant
I need to build data-schemas first:
Assistant
No TypeScript errors. Let me now run the existing tests to make sure they still pass:
Assistant
All existing tests pass. Now let me also write a test for the new fix to confirm it works. But wait, the instructions say I shouldn't modify existing tests or add new tests. Let me think about this.
Actually, the instructions say:
However UNDER NO CIRCUMSTANCES should you modify existing tests or add new tests to the repository.
So I should not add new tests. Let me instead create a temporary script to verify the fix works:
Assistant
All test cases pass. Now let me check the typescript opens more carefully. Let me also check if there are tests that exercise the exact concat logic:
Assistant
Let me run all the tests in the agents folder to ensure my fix doesn't break anything:
Assistant
All tests pass. Let me also run the Anthropic endpoint tests:
Assistant
All Anthropic tests pass. Let me also run the openai config tests:
Assistant
All tests pass. Let me also run the entire test suite for the api package to be safe:
Assistant
Let me check if these failures are caused by my fix or pre-existing:
Assistant
These failures look like infrastructure issues (Redis, MongoDB) - not related to my fix. Let me confirm by running tests on git stash:
Assistant
Yes, the failures are pre-existing - they exist even without my fix. Let me restore my fix:
Assistant
Wait, there are modifications to package.json and package-lock.json too. Let me check what changed:
Assistant
Now let me confirm the fix is still in place:
Assistant
The fix is in place. Now let me re-run the tests to make sure they still pass:
Assistant
All 343 tests pass. Let me also verify that the type compiles cleanly:
Assistant
Let me also check the linting:
Assistant
No lint errors. Now let me also think about whether the fix might miss any edge cases:
- Anthropic with two web_search tools: ✅ Fixed
- Anthropic with only LibreChat web_search (no LLM config): ✅
options.toolsis empty, somergedOptionsToolsis empty,tools = structuredToolsonly - Anthropic with only LLM web_search (no LibreChat): ✅
structuredToolsis empty,lcToolNamesis empty set,mergedOptionsTools = options.tools,tools = options.tools - OpenAI with web_search (no name): ✅ Filter handles tools without
namecorrectly - Provider with tools that have no name: ✅ Filter handles them correctly
- Provider with multiple conflicting tools: ✅ Filter removes all conflicting ones
- Provider with mixed tools (some named, some not): ✅ Filter handles correctly
Let me also write a test case to verify the actual fix works in the context of the test mocks:
Assistant
The TS enum doesn't run directly. Let me simplify my test:
Assistant
The fix works as expected. Now let me also think about what happens elsewhere. Let me look at the related code in
@librechat/agentsthat might also need updating:Assistant
Good - no other places in the agents package do this concat. The fix is minimal and localized.
Let me also think about whether the fix might affect the
toolDefinitionsarray. Looking at the code:The
toolDefinitionsis used only vialoadTools(which usesagent.toolsas input), not fromoptions.tools. Theoptions.toolsis only added to theagent.tools(i.e.,toolsarray) via the concat at line 846.So the fix only needs to be in the
toolsarray, not intoolDefinitions.Let me also check how the bindTools process works to make sure the fix doesn't break anything:
Assistant
Good. The
toolsarray (which is fixed) is what's used for binding.Now let me also make sure my fix doesn't break the case where
options.toolscould beArray<unknown>:Looking at the code:
let tools: GenericTool[] = options.tools?.length ? (options.tools as GenericTool[]) : (structuredTools ?? []);The
options.toolsis typed asunknown[]inInitializeResultBase. So the castas GenericTool[]is already happening. My fix correctly handles this withas unknown[].Let me also verify the test runs successfully one more time:
Assistant
Now let me also do a check to ensure the fix is minimal and clean:
Assistant
The fix is clean and minimal. Let me also verify by checking the final state of the file:
Assistant
Now let me cleanup my temporary files:
Assistant
Let me also do a final validation by running the broader test suite once more:
Assistant
All 516 tests pass. Let me also do the final TypeScript check:
Assistant
Now let me also lint the modified file:
Assistant
All checks pass. Let me also verify the fix doesn't change anything in the final test for the related tests:
Assistant
All tests pass. Let me also check the test file for the agent initialize to see the test for the existing concat logic:
Assistant
The existing tests for the GOOGLE_TOOL_CONFLICT logic still work. My fix is backward-compatible.
Now let me check the final diff:
Assistant
Only the file I intended to modify is changed. Let me view the final diff:
Assistant
The fix is clean and minimal. Let me now go through the checklist:
Checklist:
-
If an error was reported in the issue description, I have successfully reproduced it.
- ✅ Yes. I created a reproduction script that confirmed the issue: when both LibreChat's
web_searchand Anthropic'sweb_search_20250305are enabled, both tools get concatenated withname: 'web_search', causing the duplicate.
- ✅ Yes. I created a reproduction script that confirmed the issue: when both LibreChat's
-
If an error was reported in the issue description, I have confirmed that my fix resolves the error.
- ✅ Yes. After applying the fix, the reproduction script shows only one
web_searchtool is now exposed (LibreChat's, which is preferred).
- ✅ Yes. After applying the fix, the reproduction script shows only one
-
I have thought about edge cases and made sure my fix handles them as well.
- ✅ Yes. I verified:
- Anthropic with both web_search: Only LibreChat's is exposed (preferred)
- Anthropic with only LibreChat web_search: works (no changes)
- Anthropic with only LLM web_search: works (no changes)
- Anthropic with non-conflicting tools: works (no changes)
- OpenAI tools without name: passes through correctly (no name collision)
- Tools with various shapes: properly handled via runtime type guards
- Tools with non-string names: properly handled
- Null/undefined tools: properly handled
- ✅ Yes. I verified:
-
I have run existing tests in the repository that might have been affected by the change I made and confirmed that they pass.
- ✅ Yes. I ran:
src/agents/__tests__/initialize.test.ts(34 tests pass)src/agents/__tests__/load.spec.ts(passes)src/agents/__tests__/skills.test.ts(passes)src/agents/__tests__/handlers.spec.ts(40 tests pass)src/agents/discovery.spec.ts(passes)src/endpoints/anthropic(91 tests pass)src/endpoints/openai(218 tests pass)- Total: 516 tests pass
- I also verified the previous failures (17 test suites, 78 tests) are pre-existing and not related to my fix.
- ✅ Yes. I ran:
-
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.
Identical first prompt
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-27BCompare side by side →
Qwen/Qwen3.6-27BCompare side by side →