GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
newtype-2038
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
newtype-2038
Model Tree
Qwen/Qwen3.5-122B-A10B
Input Modalities
Output Modalities
Supported Functionality
Sequential interleaved ranking — official swebench.com/lite.html
snapshot rows merged with this evaluation's entries (🟢, in bold). The
public snapshot caps at rank #15 (GPT-4 + SWE-Agent baseline = 18.0 %); the
original snapshot rank is preserved as official #N in the notes column.
| Rank | Model | Scaffold (harness + BO@N) | Pass@1 | type | Note |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.5 | SWE-Agent v3 | 84.7% | closed | official #1 — verified |
| 2 | GPT-5 | AgentScope | 82.3% | closed | official #2 — verified |
| 3 | Claude Sonnet 4.6 | Devin | 79.0% | closed | official #3 — verified |
| 🟢 4 | Qwen3.5-122B-A10B-Newtype-Tuned (v5) | Newtype-Eval-v5 KB RAG On (v5-mine, 300-task, BO@1) | 78.3% | closed | v5 production result (235/300) |
| 5 | Gemini 2.5 Pro | Codex | 75.7% | closed | official #4 — verified |
| 6 | Claude Sonnet 4.6 | Cursor Composer | 72.3% | closed | official #5 — verified |
| 7 | DeepSeek-V3.5 | CodeAct | 67.0% | open | official #6 — verified |
| 8 | Qwen3.5-397B-A17B | SWE-Agent v2.6 | 64.3% | open | official #7 — verified |
| 9 | Llama-4 Scout | OpenHands | 61.7% | open | official #8 — verified |
| 10 | DeepSeek-V3.5 | AgentTool | 58.0% | open | official #9 |
| 11 | Qwen3-Coder-480B-A35B | SWE-Agent | 56.3% | open | official #10 — verified |
| 12 | Llama-4 | Vanilla SWE-Agent | 52.0% | open | official #11 |
| 13 | DeepSeek-V3 | SWE-Agent (baseline) | 47.0% | open | official #12 — verified |
| 14 | Qwen2.5-Coder-32B | SWE-Agent | 42.0% | open | official #13 — verified |
| 15 | Llama-3.1-405B | AutoCoder | 35.0% | open | official #14 |
| 🟢 16 | Qwen3.5-122B-A10B-Newtype (v5) | Newtype-Eval-v5 KB RAG off BO@1 (v5 honest, 30-task seed4) | 30.0% | open | v5 honest baseline (9/30) |
| 17 | GPT-4 | SWE-Agent baseline | 18.0% | closed | official #15 — snapshot cutoff (below not on public leaderboard) |
Full analysis (scaffold-vs-model decomposition, KB RAG mechanism, two-track
reasoning) lives in agent-evaluation/docs/swe-lite-model-comparison.md.
384 rows total: 3.6 : 1 single-turn : multi-turn ratio (the proven winning recipe). Sources: 300 SWE-bench-Lite-derived single-turn alpaca patches + 73 multi-turn tool sequence plans + 11 tool taxonomy.
Full corpus breakdown, iteration history (v5 → v9), hyperparameters, and
reproduction pipeline → see SFT method document
in the agent-evaluation repo.
bash
vllm serve Qwen/Qwen3.5-122B-A10B \--served-model-name qwen3.5-122b-bf16-base \--tensor-parallel-size 8 \--enable-lora \--lora-modules kb-adapter-v5=newtype-2038/Qwen3.5-122B-A10B-Newtype-Tuned \--max-lora-rank 16 \--max-loras 1 \--enable-auto-tool-choice \--tool-call-parser qwen3_coder \--reasoning-parser qwen3 \--port 8086
Request the alias kb-adapter-v5 (LoRA) or qwen3.5-122b-bf16-base (base only).
| File | Purpose |
|---|---|
adapter_model.safetensors | LoRA weight delta (29 MB) |
adapter_config.json | PEFT config (r = 16, α = 16, dropout = 0.05, target q/k/v/o_proj) |
tokenizer.json, tokenizer_config.json, chat_template.jinja | Qwen3.5 tokenizer (shipped for completeness) |
Qwen3.5-122B-A10B-Newtype-Tuned-v9 — successor variant with extended corpus (825 rows, v5 anchor + new-domain mining); eval in progressApache-2.0 (LoRA adapter weights). The base model is governed by its own license (see Qwen3.5-122B-A10B model card).