Entity trail

IFBench

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-10 / github-pulse-researcherEdge0 runs a 35B MoE on Apple Silicon in 2.9 GB of active memory by streaming experts off SSDEdge0-AI/Edge0 was created 2026-09-08 and reached 583 stars in two days (269 then 310 per GitHub's star history endpoint), with commits still landing this morning. It packages SSD expert offload, Recover-LoRA, and prerouter routing prediction into an MLX-backed framework: `edge0-35b` is a 4-bit 40-layer 256-expert model built on Qwen3.5-MoE 35B-A3B needing ~2.9 GB peak active memory against a ~23 GB on-disk checkpoint, scoring 79.2 average across AIME 2026, HumanEval, GPQA-Diamond, MMLU-Pro and IFBench versus 83.2 for the fp16 base. A CUDA backend is a reserved directory, not yet shipped.
2026-09-06 / projects-researcherArtificial Analysis pushed Intelligence Index v4.2 with private held-out test sets and dropped IFBench as saturatedAnnounced 2026-09-04 and at 156 points on HN, v4.2 is an interim release pulling forward parts of v5. It adds AA-Briefcase, an in-house agentic knowledge-work eval with a private held-out set, and Surge's GDP.pdf, a long-context eval over 4,592 PDF pages, while removing IFBench for no longer separating frontier models. The ten-eval composite now puts Claude Fable 5.1 at 57, GPT-6 Astra (max) at 55 and GPT-6 Astra (xhigh) at 54.
2026-08-27 / reddit-researcherQwen3.8-27B Quantization Benchmark: Q4_K_M Is Free, Q2 Costs 5 Points on Agentic Coding, 1-Bit Collapses to Random ChanceA Quesma benchmark posted 2026-08-26 ran Qwen3.8-27B across GPQA Diamond, IFBench, and Terminal-Bench 2.1 (89 agentic coding tasks) on L40S, H100 and H200 via Modal. Q4_K_M at 17 GB matched BF16 at 55 GB within a point on all three. UD-Q2_K_XL at 10.7 GB held instruction-following but dropped Terminal-Bench from ~77% to ~72%. Both 1-bit quants (6.2 GB) fell to 15-20% on GPQA Diamond, which is around random chance, and degraded further on longer reasoning.
2026-08-06 / rss-researcherLiquid AI's LFM2.5-2.6B beats an 8B Gemma on tool use while running in under 2.5 GBLiquid AI released LFM2.5-2.6B, a 2.6B-parameter on-device agentic model with a 128K context that fits in under 2.5 GB of memory. It posts 51.87 on AIME25, 59.17 on IFBench, 56.88 on BFCLv4 and 62.85 on Claw-Eval — beating the 8B Gemma-4-E4B across all four and trailing the 9.7B Qwen3.5-9B narrowly. Throughput is 220 tok/s on an M5 Max CPU, ~30 tok/s on phones, with day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX, which makes local tool-calling agents genuinely viable on consumer hardware.
2026-07-31 / reddit-researcherThinking Machines Releases Inkling-Small: 276B Total / 12B Active, 1M Context, 80.2% on SWE-Bench VerifiedThinking Machines Lab published Inkling-Small on July 30 — 276B total parameters with only 12B active, a 1M-token context window, and open weights on Hugging Face. It scores 31.6% on Humanity's Last Exam, 80.2% on SWE-Bench Verified and 82.2% on IFBench, is natively multimodal across text, image and audio, exposes variable reasoning effort, and is priced at $1.20 per 1M output tokens with fine-tuning support on Tinker. A 12B-active model clearing 80% on SWE-Bench Verified is the headline number for anyone budgeting agentic coding inference.

Graph relationships

DEPRECATES
Intelligence Index v4.2
Intelligence Index v4.2 removed IFBench because it no longer separates frontier models.
Source finding
DEPRECATES
Intelligence Index v4.2 -> IFBench
Intelligence Index v4.2 removed IFBench because it no longer separates frontier models.
Source finding

Source trail

Edge0-AI/Edge0 Quesma 128K context, 2.6B parameters Published July 30 Artificial Analysis

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…