Cogferra · Benchmark Front End

Benchmark intelligence
for agent-grade models

Benchmark runs for local models - decode speed, concurrency, and long-context scaling - collected in one place and measured the same way, so the numbers actually compare.

Browse results Model pages

Corpus BetterBench · 15 categories

Harness IronLLM Test Harness

Serving vLLM fork · gfx1201

6Model families tracked with per-build result sets

150+Published benchmark reports across the build line

128KContext depth exercised by the scaling sweep

16Concurrency levels per run, batch-swept end to end

01 · Results

View and upload results

Browse a single run, compare across runs, or drop in a new one.

Viewer

Charts and tables for a single run: decode, ITL, TTFT, concurrency and prefill sweeps.

Compare

Overlay two or more runs to isolate what a kernel, flag or quantization actually changed.

Details

Per-category breakdowns, variance and per-run configuration at full precision.

Upload

Ingest a new BetterBench result set into the store - cross-checked against the published report.

02 · Models

The models we've put through it

Each model gets its own page with its full result history.

QWEN

Qwen3.x-27B

Dense/moe 27B line - RFA/RFI quant ablations and Dflash drafts.

QWEN

Qwen3.8-Flash-Next

Hybrid MoE flagship - MTP speculative decode, HC fusion, the Rust kernel line.

MUSE

Muse-Glimmer

30B-class long-context runs with MTP drafts at 512K.

GEMMA

Gemma4-31B

Single-family results as they enter the bench.

GLM

GLM-5.3-Flash

Single-family results as they enter the bench.

DEEPSEEK

DeepSeekV4-Family

Vision-flash and family members as results land.

03 · Reference

Reading between the numbers

The methodology behind the corpus and the engineering behind the serving stack.

vLLM serving docsFlags, kernels & build line Effective Context EngineeringField guide · why benchmarks look like this

Benchmark intelligencefor agent-grade models