Connect Mellum 2.1 by running it behind an OpenAI-compatible chat-completions server, then point your coding agent at that server’s base URL and model ID. For the original weights, JetBrains documents vLLM and SGLang. For a smaller local deployment, its official GGUF repository documents llama.cpp and setup examples for Pi, Hermes Agent, and OpenClaw. Tool-using agents need both a compatible endpoint and the backend’s tool-call settings.
What the connection actually involves
An existing coding agent normally does not load model weights itself. It sends an HTTP request containing messages, optional tools, and generation settings to a model endpoint. The server running Mellum 2.1 returns an assistant message, a reasoning segment where supported, or a structured tool call. Your integration therefore has four pieces:
As an Amazon Associate I earn from qualifying purchases.
- Model files: the original Mellum 2.1 weights or a GGUF quantization.
- Inference server: vLLM, SGLang, llama.cpp, or another runtime that exposes an OpenAI-compatible API.
- Endpoint settings: the server’s base URL, usually ending in
/v1, and the exact model identifier it advertises. - Agent settings: a custom provider or client that sends chat-completions requests and understands the returned format, including tool calls if your agent uses shell, file, or other tools.
Mellum 2.1 is a 12-billion-parameter mixture-of-experts model with 2.5 billion active parameters, and its published context length is 131,072 tokens. Those are model specifications, not promises that any particular computer can hold the full context or generate at a useful speed. The official references are the JetBrains model card and the JetBrains GGUF repository.
Decide which serving route fits your agent
| Route | What you run | Best fit | Important qualification |
|---|---|---|---|
| vLLM | Original Mellum 2.1 weights | A GPU server needing OpenAI-compatible chat completions and documented tool calling | JetBrains’ tool example uses specific reasoning and Hermes parser flags; copy those settings rather than assuming defaults. |
| SGLang | Original Mellum 2.1 weights | A deployment already standardized on SGLang | The documented example uses port 30000 and /v1/chat/completions; the model card does not provide a Mellum-specific tool-parser configuration for SGLang. |
| llama.cpp | A GGUF quantization | Local inference, a compact workstation, or a client that only needs an OpenAI-compatible local endpoint | Memory, speed, and context capacity depend on the selected quantization, hardware, and context in use. |
| Another OpenAI-compatible client or SDK | Any server exposing the expected API | An agent with a custom provider or an application using an OpenAI-compatible client | The client still needs the exact model ID and a request/response format it can parse. |
Choose the runtime before editing the agent. The runtime determines the URL, model-name convention, authentication behavior, supported tool schema, and resource requirements.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Prepare the agent and server settings
1. Find the endpoint shape your agent expects
Some agents ask for a base URL, such as http://localhost:8000/v1, and append /chat/completions themselves. Others ask for the complete path, such as http://localhost:8000/v1/chat/completions. Supplying the full path where a base URL is expected can produce a duplicated path and a 404 response. Read the selected agent’s provider documentation and enter exactly the form it requests.
2. Record the served model identifier
The identifier in the agent must match the identifier accepted by the running server. With original weights it may be the repository name. With GGUF, the launch command can include a quantization suffix while a client example uses the repository’s model identifier. Do not “simplify” the name; inspect the runtime’s advertised model list or copy the convention from its current documentation.
3. Decide whether an API key is needed
A local server may accept any placeholder key or may be configured to require a real value. Enter the value required by that server. A key accepted by the client does not make an unauthenticated local endpoint secure, so bind the server appropriately and do not expose it to an untrusted network without authentication and access controls.
Free tools Windows power users keep installed
One-click scans. No signup required.
4. Check tool support before enabling tools
A plain text completion proves only that generation works. A coding agent also needs the server to accept its tool definitions and return calls in the schema the agent understands. Test a harmless tool request before giving the model access to a repository or shell.
Option 1: serve the original weights with vLLM
JetBrains’ model card provides the clearest documented route for an OpenAI-compatible server with tool calling. Install a current vLLM build according to its own documentation, then start Mellum with:
vllm serve JetBrains/Mellum2.1-12B-A2.5B-Thinking
--max-model-len 131072
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser hermesThe --max-model-len 131072 value reflects the model card’s published context length. It does not guarantee that your hardware can sustain that context. The reasoning parser tells vLLM how to separate the model’s reasoning output, while automatic tool choice and the Hermes parser are the documented settings for tool calls.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Connect the agent to vLLM
- Start the server and wait until it reports that it is listening.
- Open your agent’s custom provider, OpenAI-compatible provider, or local-model settings.
- Set the base URL to
http://localhost:8000/v1, unless the agent specifically asks for the full/v1/chat/completionsURL. - Set the model field to the exact model ID exposed by vLLM.
- Enter an API key only if your vLLM deployment requires one.
- Send a short coding request without tools, then run a second request that invokes one safe, read-only tool.
If text generation succeeds but tool calls arrive as ordinary prose, verify all three layers: the vLLM flags, the agent’s tool schema, and the parser expected by the agent. The vLLM command above is documented for this model; its flags should not be assumed to transfer unchanged to a different backend.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Reference: JetBrains’ Mellum 2.1 model card.
Option 2: serve Mellum with SGLang
JetBrains also documents launching the original weights with SGLang on port 30000 and sending requests to the OpenAI-compatible route /v1/chat/completions. Use the current SGLang launch syntax for your installed version, then configure the agent with the resulting base URL, normally http://localhost:30000/v1 when the agent appends the completion path.
- Launch the model with SGLang using the model identifier from the model card.
- Confirm that the server is listening on port 30000 and that its OpenAI-compatible route responds.
- In the agent, select its custom OpenAI-compatible provider.
- Enter
http://localhost:30000/v1as the base URL, or the full route only if the agent requests it. - Use the exact model ID reported by SGLang.
- Test a normal completion before testing tools.
The model card does not give a Mellum-specific SGLang tool-call parser equivalent to vLLM’s --tool-call-parser hermes setting. If your agent depends on tools, verify the current SGLang tool-calling implementation and the agent’s expected schema instead of assuming that a successful text response proves compatibility.
Reference: JetBrains’ Mellum 2.1 model card.
Option 3: run a GGUF quantization with llama.cpp
For local deployment, the official GGUF repository supplies quantized files and a llama.cpp server command. Its Q4_K_M variant is described by the repository as a balanced, recommended choice; that is the publisher’s characterization, not an independent benchmark conclusion.
llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M
--ctx-size 131072 --temp 0.6 --top-p 0.95 --top-k 20This starts an OpenAI-compatible local endpoint at http://localhost:8080/v1 in the repository’s example. The --ctx-size value requests the published context length, while --temp, --top-p, and --top-k are the example’s sampling settings. Lowering the context can reduce memory pressure, but it changes how much repository or conversation history fits in one request.
Available file sizes
| Quantization | Published file size | How to interpret it |
|---|---|---|
| MXFP4_MOE | 7.0 GB | Smallest listed file; file size alone is not a minimum-RAM or minimum-VRAM specification. |
| Q4_K_M | 8.1 GB | The GGUF repository labels this the balanced, recommended example. |
| Q6_K | 10.9 GB | Larger file with a different quality/resource trade-off. |
| Q8_0 | 12.9 GB | Higher-precision quantized file than the Q4 and Q6 options listed above. |
| BF16 | 24.3 GB | Largest listed file; still not a complete system requirement. |
The repository also reports KLD and top-token-match measurements against BF16 logits on Wikitext-2 with a 512-token context. Those measurements describe quantization behavior under that test setup; they do not predict every coding task, agent, backend, or hardware configuration.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
Register the local endpoint in an agent
The GGUF repository includes setup examples for several agents. In Pi, the example registers llama-cpp in ~/.pi/agent/models.json, sets baseUrl to http://localhost:8080/v1, uses api: "openai-completions", and supplies the GGUF model identifier before starting Pi in the project directory. Hermes Agent uses a custom provider and the same local base URL pattern. OpenClaw uses a custom base URL, model ID, provider ID, and an OpenAI-compatibility setting.
Those configuration snippets are tied to the respective agent versions and can change. Open the live GGUF repository instructions, copy the current field names, and preserve the exact model-ID convention shown for your selected quantization.
Option 4: connect an OpenAI-compatible SDK or another agent
If your coding agent accepts a custom OpenAI client, the integration is the same: create a client with the local server’s base URL, then assign that client to the agent’s chat-completions model. The OpenAI Agents SDK documents this pattern with AsyncOpenAI and OpenAIChatCompletionsModel. A version-adapted example is:
from openai import AsyncOpenAI
from agents import Agent, OpenAIChatCompletionsModel
client = AsyncOpenAI(
base_url="http://localhost:8000/v1",
api_key="local-key"
)
model = OpenAIChatCompletionsModel(
model="JetBrains/Mellum2.1-12B-A2.5B-Thinking",
openai_client=client
)
agent = Agent(
name="Mellum coding agent",
instructions="Work in the repository and explain changes.",
model=model
)Use the model name and URL for your actual server, not necessarily the vLLM values in this example. The SDK documentation distinguishes a global client, a provider used for one run, and a model assigned to one agent, so choose the scope that matches your application. This is a generic OpenAI-compatible integration pattern rather than a Mellum-specific adapter. See the OpenAI Agents SDK model documentation.
Configure tool calling deliberately
Mellum 2.1 is intended for complex agentic work, but the model’s ability to generate a tool call is only one part of a working loop. Your agent must send a valid tool definition, the server must parse the model’s tool-call format, and the agent must execute the call and return the result in the next message.
- Start with one read-only tool, such as listing a directory or reading a file.
- Check the raw response if the agent displays prose instead of an invocation. Look for a structured tool-call object rather than text that merely describes a command.
- For vLLM, use the documented Qwen3 reasoning parser, automatic tool choice, and Hermes parser flags.
- For SGLang and llama.cpp, consult the current backend and agent documentation for tool-call support; do not copy vLLM flags that the backend does not implement.
- Keep shell and write access disabled until a read-only call completes and the agent returns the result correctly.
Match quantization and context to your machine
The 7.0 GB through 24.3 GB GGUF file sizes are download sizes. They do not tell you the total memory needed for model buffers, the key-value cache, the operating system, the runtime, or a 131,072-token context. Actual requirements vary with offloading, batch size, concurrent requests, context length, and backend implementation. The reviewed Mellum sources do not state a universal minimum GPU, VRAM, or RAM configuration.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
Use this decision process:
- Choose the longest context you genuinely need. A repository task that fits in 16K or 32K tokens may require substantially less memory than a request for the full 131,072-token context.
- Select a quantization whose file size and runtime overhead fit your available memory, leaving headroom for the context cache and the rest of the system.
- Start with a short prompt and measure whether generation is stable before increasing context or concurrency.
- If memory pressure appears, lower the context size, reduce concurrency, choose a smaller quantization, or move inference to a machine with more memory.
Do not buy or size a machine from the GGUF file size alone. Hardware must be matched to the quantization, runtime, context length, and workload you will actually use.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHosted availability and network planning
At the time of the official model-card review on October 9, 2026, Mellum 2.1 was not deployed by an Inference Provider, and Hugging Face’s generic compatible-router documentation did not list it as available. Hosted availability can change, so check the current model card and Inference Providers documentation before designing around a hosted endpoint. If no provider offers the model, plan for a self-hosted vLLM, SGLang, or llama.cpp server.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Verify the connection before giving the agent write access
- Health check: confirm the server process is running and listening on the expected port.
- Model check: compare the agent’s model field with the server’s advertised model ID, including any quantization suffix required by that runtime.
- Plain completion: ask for a short explanation of a small code fragment. This isolates URL and model-selection problems from tool problems.
- Context check: send a request comfortably below the selected context limit before testing a large repository prompt.
- Read-only tool check: expose one harmless tool and confirm that the server returns a structured call and that the agent sends the result back.
- Permission check: only after the previous steps succeed, enable file writes, shell commands, network access, or other consequential tools.
Troubleshoot the common failures
“Model not found” or a 404 response
Check whether the agent expects a base URL or a complete completion URL. Then compare the configured model string with the serving command and the runtime’s model list. GGUF setups are especially sensitive to repository names and quantization suffixes. Copy the exact identifier used by the selected runtime instead of translating it by hand.
The agent connects but tools are returned as text
First confirm that the agent actually included tools in the request. Next verify backend-specific tool support. In vLLM, check --enable-auto-tool-choice and --tool-call-parser hermes, together with the documented qwen3 reasoning parser. Do not assume that the same flags or parser exist in SGLang or llama.cpp. Finally, inspect whether the agent expects the tool call in a different schema.
Context-length errors or out-of-memory failures
The model card’s 131,072-token context is a published limit, not a guarantee for your deployment. Reduce the requested context, shorten repository history, lower concurrency, or use a smaller GGUF file. If you are using original weights, review the server’s memory and offloading settings. The available sources do not provide one minimum hardware profile that applies to every backend.
Slow generation
Measure with the same prompt and context after each change. Longer context, larger quantizations, CPU offload, and multiple simultaneous requests can all reduce throughput. A smaller quantization may improve responsiveness but changes the precision trade-off; the GGUF repository’s quantization measurements were made on Wikitext-2 with a 512-token context, not on your coding workload.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
The endpoint works locally but not from another machine
localhost refers to the machine running the agent, not necessarily the machine running the model server. Use the server’s reachable address only on a trusted network, configure the server to listen on the required interface, and add authentication or network controls before exposing it beyond the local host.
Sampling settings behave unexpectedly
Keep the runtime’s documented defaults while validating connectivity. The llama.cpp example sets temperature to 0.6, top-p to 0.95, and top-k to 20; the agent may send its own generation parameters and override those defaults. Make one change at a time so a tool or quality regression can be traced to a specific setting.
How to read Mellum’s published performance figures
JetBrains reports SWE-bench Verified at 47.0%, Terminal-Bench 2.1 at 17.4%, and SWE-bench Pro at 28.0%. The model card labels these evaluations self-reported. Its agentic setup used Pi v0.73.1 with shell and file tools, a 114K-token context, up to 16K generated tokens per turn, and Mellum 2.1’s default temperature of 1.0. Those figures describe that particular harness and configuration; they are not a controlled prediction for a different coding agent, backend, quantization, context, or sampling policy.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteThe card also reports an AIME 25/26 average based on 30 questions each. That is a separate vendor-reported reasoning evaluation and does not establish tool-call compatibility or repository performance in your deployment.
Recommended path for most self-hosted users
If you have a GPU setup intended for the original weights and need documented tool calling, start with vLLM and the model-card command. If you need a local endpoint with a downloadable quantization, start with llama.cpp and the repository’s Q4_K_M example, then adjust context and quantization after measuring memory and speed. Use SGLang when it fits an existing deployment standard and verify tool support separately. In every case, connect the agent only after confirming the exact base URL, model ID, and response format.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
