Research Vault

AI9Stars G9v3-39A5B

Properties1
tagshuggingface, text-generation, coding, agentic, moe, instruct, long-context, tool-calling, reasoning, custom-architecture

2 min read

AI9Stars G9v3-39A5B

One-line: 39B-total/5B-active reasoning MoE from AI9Stars with 131K context + native tool calling — AA Intelligence Index 32 (#6 in its size class), but a preview release with custom architecture and no published coding/agent-loop benchmark split yet. ⚠ surfaced via the trending window (repo lastModified 08-19; release ~08-03, no prior vault note) — not a fresh first-drop.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published
SWE-bench Multilingual—not published

No per-benchmark coding table on the card. Independent Artificial Analysis coverage: SciCode 38.2%, AA Coding Index 31.7 (disaggregated from the AA composite). No SWE-bench Verified / Aider / LiveCodeBench / BFCL figure published anywhere found — agent-loop repo-resolution suitability unverified.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published

Card carries tool-calling tag and ships supports Think / No Think modes, but publishes no BFCL / τ-Bench / ToolACE / GAIA number. Tool-use capability is claimed, not independently benchmarked.

Reasoning (context only)

BenchmarkScoreSource
AA Intelligence Index32Artificial Analysis (independent — #6/135 in class)
SciCode38.2%Artificial Analysis (independent)
GPQA Diamond~75.6% (in AA composite)Artificial Analysis (composite)
AA-LCR60.3%Artificial Analysis
MMLU / MMLU-Pro—not published
AIME—not published

Efficiency

  • Active params: 5B (MoE, 39B total — sparse design keeps inference cost near a small dense model)
  • Context length: 131,072 tokens (131K, rope_theta 5e6, no rope_scaling)
  • VRAM (fp16 rough): ~78 GB for full 39B weights (5B active). Quantization needed for consumer GPUs; no official GGUF published — custom architecture blocks stock llama.cpp loading.
  • Runs on: vLLM (≥0.21), SGLang (≥0.5.12), Transformers (≥5.6) — custom model code required (G9v3ForCausalLM, configuration_g9v3.py / modeling_g9v3.py shipped in-repo). Not loadable by stock transformers 4.x without the custom files, and not llama.cpp-compatible.

What makes it notable

G9v3-39A5B is a sparse reasoning MoE (5B active of 39B) pairing 131K context with tool-calling and Think/No-Think modes — a strong Intelligence-vs-active-params ratio for self-hosted agent use (AA Intelligence Index 32, well above the class median of 9). Treat it as a promising but preview release: the card publishes zero coding/agent-loop benchmark splits (only the AA composite + SciCode exist independently), it’s flagged preview, and its custom architecture means no llama.cpp/Ollama GGUF path yet — you’d need vLLM/SGLang or the repo’s own modeling code. For the local agent-runner it’s worth watching for a fuller eval release and community GGUF, but agent-loop repo-resolution suitability is unverified at this point. Released under Apache-2.0, not gated.

See also

Graph View