One-line: 9B agentic coding model distilled from Fable-5.1 / GLM-5.3 agent traces (Claude Code, OpenCode, Codex) — claims SWE-bench Verified 73.5 with a 262K native context.
⚠ All scores are lab-reported (OrionLLM model card, OpenHands harness, GLM-5.3 as judge for HLE/MCP-Atlas). No third-party reproduction found yet — treat as claims.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
—
—
Aider polyglot
—
—
SWE-bench Verified
73.5
lab-reported
SWE-bench Pro
49.1
lab-reported
Terminal-Bench 2.1 (Terminus-2)
49.6
lab-reported
Terminal-Bench 2.1 (Claude Code)
50.8
lab-reported
NL2Repo
36.2
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
—
τ-Bench
—
—
MCP-Atlas
56.7
lab-reported
ClawEval
67.8
lab-reported
BrowseComp
57.4
lab-reported
ToolACE
—
—
GAIA
—
—
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
86.9
lab-reported
HLE (no tools)
21.2
lab-reported
HLE (with tools)
32.8
lab-reported
MATH
—
—
AIME 2025
—
—
Efficiency
Active params: 9B total (Qwen3.5-9B base; hybrid linear+full attention as per config layer_types)
Context length: 262,144 tokens (native)
VRAM (fp16 rough): ~18 GB weights + KV cache for 262K context will exceed a single consumer GPU — quant + long-context optimizations needed for local agents
What makes it notable
Claims SWE-bench Verified 73.5 at 9B — above Ornith-1.5-9B (70.6), Ornith-1.0-9B (69.4), Qwen3.5-9B (53.2), Gemma-4-31B (52.0) on the card’s reference table.
Positioned explicitly for agent loops: trained on read-before-write patterns, LSP-diagnostic recovery, minimal-edit diffs; Terminal-Bench 2.1 49.6–50.8 is the strongest headline for CLI/terminal agent use.
Built on Qwen3.5-9B (hybrid attention, VL-capable arch) — runs locally at ~18 GB fp16; distilled from frontier agentic trajectories (Fable-5.1, GLM-5.3).
Fresh release (repo created 2026-09-07, benchmark table published 2026-09-14); no independent verification yet.