44 Minutes, 176k Tokens: Fable 5.1 Cracks a 370-Year-Old Cipher
Vals AI gave Claude Fable 5.1 Urquhart's unsolved Cyphral Distich: solved in 44 minutes, 176k tokens, zero interjections — and it went on to decode the 285-number Octastich too.
#benchmarks
23 posts
HP ZBook Ultra G1a 14: 64GB Unified Memory Hits 8.6 t/s on 32B Local Models
AMD Ryzen AI Max PRO 390 with 64GB unified memory runs 32B models at 8.6 t/s — parity with a 16GB RTX 5080 laptop, no discrete GPU needed.
GPT-6 Astra Shifts the Frontier Model Calculus—99.9% on ARC, 100% on Exploits
GPT-6 Astra saturates benchmarks and raises the bar on computer use, coding, and cybersecurity. Teams must now reconsider whether open-weights or smaller fine-tuned agents remain cost-effective.
NVIDIA's Custom HBM Signals Memory Bandwidth Is the New Constraint
NVIDIA's NVHBM moves memory control to the HBM stack itself, not the GPU. This architectural shift reveals memory bandwidth, not compute cores, is the real scaling bottleneck for AI inference.
Gemini 3.8 Flash Cuts Agent Latency—Same Speed, Bigger Reasoning Budget
Google's Gemini 3.8 Flash delivers frontier-level coding and reasoning at Flash speed and cost ($0.75/1M in tokens). A Cyber variant targets vulnerability discovery and patching.
Your 4-Bit Model Can Now Beat Its 16-Bit Original
Quantization-Aware Healing distills a 4-bit model from the pre-compression teacher, winning 7 of 9 benchmarks over its bf16 source.
Granite 4.2 Brings Agentic RL Down to a 3B Open-Weights Model
IBM's Granite 4.2 — 3B/8B/30B, Apache 2.0, 512K context, thinking/non-thinking switch, and agentic RL on 8B/30B with real SWE, terminal, and search environments.
GLM-5.3 Hits 88.2 on Terminal Bench 2.1 as Open-Weights
Z.ai open-sourced GLM-5.3 (753B MoE). 88.2 on Terminal Bench 2.1 and 28.3 on TB 3.0, serving via vLLM/SGLang with a reasoning_effort flag and 23 existing quantizations.
DeepSeek V4 Flash 0731: 82.7 on Terminal Bench, at a Third of Pro's Output Price
DeepSeek's re-post-trained DeepSeek-V4-Flash-0731 scores 82.7 on Terminal Bench 2.1 — reportedly beating its own near-6x larger V4-Pro-Preview — and now speaks the Responses API natively for Codex-style harnesses. What actually shipped, how to split high vs max effort, and the verbosity caveat that quietly resets the per-task math.
AMD EPYC Venice Launches July 22–24: 576 Cores, 12 TB/s Memory Bandwidth
AMD's Zen 6 EPYC Venice debuts next week with up to 576 cores, 12 TB/s memory bandwidth, and native AVX-1024 support. Here's what changes for AI infrastructure.
Stacking RL on Already-RL Models Breaks the Ceiling
Cognition's SWE-1.7 adds 12.2 points atop an already-trained base, challenging the field's assumption that RL post-training hits a hard wall.
295B Parameters, 21B Active—Hy3 Proves Size Isn't Agent Performance
Tencent's Hy3 benchmarks between models 2–5x larger by parameter count. 256K context, Apache 2.0, 1 yuan per million input tokens. The catch: vendor benchmarks need independent testing.
GLM-5.2 Shifts Agentic Coding Economics to 1/6th Frontier Cost
Z.AI's open-weights GLM-5.2 hits 81% on Terminal-Bench, undercutting GPT-5.5 for a fraction of the cost. Here's the benchmark breakdown.
DiffusionGemma Puts 1000+ tok/s on an RTX 5090
Google's open 26B diffusion model hits 700+ tok/s on consumer GPUs with day-zero vLLM support. Here's what the bidirectional architecture changes for local inference.
Apple Silicon Just Cracked the 150B Wall at 56 tok/s
Rapid-MLX's new engine optimizations push DeepSeek V4 Flash 158B-A13B to 56 tok/s on a Mac. Here are the exact numbers and the tradeoffs.
Your 12GB MTP Throughput Just Jumped 23%
A community llama.cpp fork squeezes 110 tok/s out of Qwen3.6-35B-A3B MTP on a 12GB card. Here are the exact flags and the VRAM trick to make it fit.
Your Local Throughput Just Doubled (With No Accuracy Tax)
Multi-token prediction is merged into llama.cpp, delivering 1.4–2.2x throughput on Qwen3.6 with zero accuracy loss.
MTP in llama.cpp Adds 24–33% Throughput on Qwen3
llama.cpp merged Multi-Token Prediction for Qwen3. Community benchmarks show 38→47 tok/s on RTX 3090 and 63→84 tok/s on RTX 5090 — no new hardware needed.
DFlash Decode Collapses at Long Output. MTP Doesn't.
One llama.cpp flag—--spec-draft-p-min 0.75—turns MTP from a dud into a decode speed that holds flat across output lengths where DFlash falls apart.
One llama.cpp Flag Turns MTP From Dead Weight to 68% Faster
The --spec-draft-p-min filter in llama.cpp PR #22397 rescues MTP for Qwen3.6-27B: 48.9 tok/s vs 29 tok/s at 2000 tokens on a 24GB card.
Your vLLM Thinking Budget Was Doing Nothing With MTP On
vLLM 0.21.0 shipped Friday with a quiet fix: thinking_token_budget was being silently ignored when MTP speculative decoding was enabled. If you serve reasoning models with spec decode, you have been paying for it.
Claude Code v2.1.100+ Burns ~20K Phantom Tokens Per Request
A server-side bug in Claude Code v2.1.100+ inflates every request by roughly 20K cache_creation tokens — about 40% overhead. Pin v2.1.98 until fixed.
A 27B Model on a Single GPU Is 10 Points Off Claude Opus 4.7
Qwen3.6-27B running locally now scores within 10 points of frontier closed models on SWE-bench Verified. The benchmark table, lined up side by side.