Model Lineup

Weaver supports multiple open-weight base models and lets the server choose training resources based on model, training mode, and performance tier.

TIP

Available models can change with server configuration. Use service_client.list_supported_models() to query currently healthy models.

Querying Models

from weaver import ServiceClient

with ServiceClient() as service_client:
    models = service_client.list_supported_models()
    print(models)

Get server configuration for a model:

config = service_client.get_supported_model_config("Qwen/Qwen3-8B")
print(config)

Common Models

Nex-AGI Optimized Models

These models are optimized for reasoning, instruction following, tool use, and agent workflows.

Model ID	Parameters	Type	Context Length
nex-agi/Qwen3-30B-A3B-Nex-N1	30B, about 3B active	MoE	128K
nex-agi/Qwen3-32B-Nex-N1	32B	Dense	128K
nex-agi/DeepSeek-V3.1-Nex-N1	671B, about 37B active	MoE	128K

Qwen Series

Model ID	Parameters	Type	Context Length
Qwen/Qwen3-8B	8B	Dense	128K
Qwen/Qwen3-32B	32B	Dense	128K
Qwen/Qwen3-30B-A3B	30B, about 3B active	MoE	128K
Qwen/Qwen3-235B-A22B	235B, about 22B active	MoE	128K

DeepSeek Series

Model ID	Parameters	Type	Context Length
deepseek-ai/DeepSeek-V3.1	671B, about 37B active	MoE	128K
deepseek-ai/DeepSeek-V3.2	671B, about 37B active	MoE	128K

Training Modes

LoRA

The default mode. It is a good fit for fast experiments, lower-cost fine-tuning, and frequent RL updates.

training_client = service_client.create_model(
    base_model="Qwen/Qwen3-8B",
    training_mode="lora",
    lora_config=types.LoraConfig(
        rank=32,
        train_attn=True,
        train_mlp=True,
        train_unembed=True,
    ),
)

Full Fine-Tuning

Full fine-tuning is useful when LoRA capacity is insufficient or when the target behavior requires deeper model adaptation.

training_client = service_client.create_model(
    base_model="Qwen/Qwen3-8B",
    training_mode="full_ft",
)

Long Context Models

Some models support maximum context length suffixes in base_model:

training_client = service_client.create_model(
    base_model="Qwen/Qwen3-8B:262144",
    training_mode="full_ft",
)

When using a long-context variant, check data length, training cost, and throughput expectations.

Performance Tiers

Use performance_tier to request different throughput tiers:

training_client = service_client.create_model(
    base_model="Qwen/Qwen3-8B",
    performance_tier="fast",
)

Common values:

• normal: default or standard throughput.
• fast: higher throughput, usually at higher cost.
• flash: higher-performance tier, subject to server availability.

Selection Guidance

First Weaver run: start with Qwen/Qwen3-8B, default LoRA, rank 32.

SFT baseline: start with 8B or 32B dense models to debug data format, masks, and learning rate.

Agent scenarios: evaluate nex-agi/* optimized models first, especially for tool use and multi-step reasoning.

High-quality RL: use larger Qwen or DeepSeek MoE models when budget allows, and let NexRL manage rollout and policy updates.

Long-context tasks: use an explicit context suffix and control token counts in sampling, rewards, and training batches.

Next Steps

• Training and Sampling: create models and start training.
• Loss Functions: choose the objective.
• Saving and Loading: save training state and sampler weights.