GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AllSpark-Research
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AllSpark-Research
Model Tree
Qwen/Qwen3.6-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
Navalia-35B-A3B is an open-weight cowork agent post-trained from Qwen3.6-35B-A3B. It is designed for tasks that combine document retrieval, cross-document evidence integration, numerical analysis, and tool use.
The Navalia framework connects verifiable task synthesis with long-horizon agent post-training. Starting from public corporate financial reports, it selects supporting evidence and builds executable computation chains to construct analytical tasks with traceable sources and verifiable answers. These tasks provide a foundation for training agents to investigate documents, perform calculations, and carry an analysis through multiple tool interactions.
| Base | Qwen3.6-35B-A3B |
|---|---|
| Parameters | 35B total / 3B active (256 experts, 8 active) |
| Layers / hidden | 40 / 2048 |
| Context | 256K |
| Precision | bfloat16 |
Results reported in the Navalia paper on financial research and professional workplace tasks. Navalia improves over Qwen3.6-35B-A3B on all five benchmarks. Scores are percentages; parentheses in the final row show gains over the base model in percentage points.
| Model | OfficeQA ProCorrectness ↑ | FinSearchCompT2&T3 Acc. ↑ | Finance AgentPublic-77 CB All-Pass ↑ | APEX-Agents-AAPass@1 ↑ | Workspace-Bench-LiteRubric Pass Rate ↑ |
|---|---|---|---|---|---|
| Proprietary Frontier Models | |||||
| GPT-5.4* | 50.40 | – | – | 33.26 | 52.9 |
| Claude Opus 4.6* |
* Results taken from the original papers, technical reports, or official websites; other results were obtained under the paper's unified evaluation setup. A marker on a model name applies to the entire row; a marker on a score applies only to that entry. † Results use a different document representation or agent harness and are included for reference.
Navalia uses a lightweight ReAct harness for trajectory collection and online RL rollouts: the agent reasons about the task, calls tools, reads their outputs, and continues until it can produce a final answer. Tools cover shell execution, file reading and editing, document search, and Serper web search. Our experiments run each task in an isolated sandbox, using the original system prompt and accounting conventions. The public harness includes tool definitions and a generic execution interface; connect your own sandbox backend to run trajectory collection.
Long trajectories use FIFO context management. The system prompt and original question stay in context. When the history exceeds 30 assistant turns, the oldest 10 turns are removed together with their tool calls and results. Working files remain available throughout the task, and saved trajectories retain the full interaction history.
The harness preserves reasoning from previous turns, uses native function calling, and recognizes
<FINAL_ANSWER>...</FINAL_ANSWER> as the completion marker. The default interaction budget is
200 turns, with up to 65,536 generated tokens per response.
Serve the model with SGLang. The example below uses four GPUs:
bash
pip install 'sglang[all]>=0.5.10' python -m sglang.launch_server \ --model-path AllSpark-Research/Navalia-35B-A3B \ --served-model-name Navalia-35B-A3B \ --port 8000 \ --tp-size 4 \ --context-length 262144 \ --reasoning-parser qwen3 \ --tool-call-parser qwen3_coderAdjust tensor parallelism and context length to your GPU memory. The server exposes an
OpenAI-compatible API at http://localhost:8000/v1.
Send a request:
bash
curl http://localhost:8000/v1/chat/completions \ -H 'Content-Type: application/json' \ -d '{ "model": "Navalia-35B-A3B", "messages": [ { "role": "user", "content": "Revenue rose from $120 million to $150 million. Calculate the growth rate and explain the calculation." } ], "temperature": 1.0, "top_p": 0.95, "max_tokens": 8192, "chat_template_kwargs": { "enable_thinking": true, "preserve_thinking": true } }'For agent tasks, provide tool definitions through the API's tools field, execute returned tool
calls in your environment, and append their results to the conversation. Keep each assistant
message's reasoning_content when continuing the conversation so that the chat template can
preserve its reasoning.
Apache 2.0.
pip install 'sglang[all]>=0.5.10' python -m sglang.launch_server \ --model-path AllSpark-Research/Navalia-35B-A3B \ --served-model-name Navalia-35B-A3B \ --port 8000 \ --tp-size 4 \ --context-length 262144 \ --reasoning-parser qwen3 \ --tool-call-parser qwen3_codercurl http://localhost:8000/v1/chat/completions \ -H 'Content-Type: application/json' \ -d '{ "model": "Navalia-35B-A3B", "messages": [ { "role": "user", "content": "Revenue rose from $120 million to $150 million. Calculate the growth rate and explain the calculation." } ], "temperature": 1.0, "top_p": 0.95, "max_tokens": 8192, "chat_template_kwargs": { "enable_thinking": true, "preserve_thinking": true } }'| 55.60 |
| – |
| – |
| 33.04 |
| – |
| Claude Opus 4.7* | 46.6† | – | – | – | 66.6 |
| Gemini 3.1 Pro Preview* | 39.10 | – | – | 32.01 | 38.7 |
| Seed2.1 Pro* | 70.9† | – | – | – | 53.0† |
| Qwen3.7 Plus | 55.39 | 61.1 | 56.94 | 22.42* | 55.01 |
| Large Open-Weight Models |
| Kimi-K3 | 67.92 | 68.8 | 65.02 | 37.13 | 60.56 |
| DeepSeek-V4 Flash | 65.66 | 65.2 | 58.87 | 36.43 | 58.08 |
| Qwen3.8 Flash Next | 64.04 | 70.3 | 58.48 | 38.86 | 61.96 |
| GLM-5.2 | 63.41 | 65.7 | 63.1 | 33.89 | 59.34 |
| Kimi-K2.6 | 58.65 | 59.1 | 54.49 | 29.28 | 54.96 |
| Qwen3.5-397B-A17B | 47.62 | 50.4 | 51.82 | 14.60 | 50.51 |
| Baseline |
| Qwen3.6-35B-A3B | 39.6 | 41.7 | 43.19 | 12.24 | 46.2 |
| Navalia-35B-A3B | 62.66 (+23.06) | 55.80 (+14.10) | 57.10 (+13.91) | 25.44 (+13.20) | 59.27 (+13.07) |