{ "version": "https://jsonfeed.org/version/1.1", "title": "Frontier Checkpoint", "home_page_url": "https://frontiercheckpoint.com/", "feed_url": "https://frontiercheckpoint.com/feed.json", "description": "A practitioner-only technical publication for working ML and agent engineers. We verify, reproduce, and recreate the work — exploring the open frontier of what actually holds up.", "language": "en-US", "icon": "https://frontiercheckpoint.com/favicon-32.png", "authors": [ { "name": "Frontier Checkpoint", "url": "https://frontiercheckpoint.com" } ], "items": [ { "id": "https://frontiercheckpoint.com/essays/what-a-4b-model-can-do/", "url": "https://frontiercheckpoint.com/essays/what-a-4b-model-can-do/", "title": "What a 4B Model Can Actually Do: Field Notes from 155 Experiments", "summary": "Across 155 small-model experiments centered on Qwen 3.5 4B, the same thing kept working: give the model something executable it can check against the evidence it has, and it punches far above its benchmark weight. Here is the field guide — the levers that worked, how I know they're real, and the frontier they opened up.", "content_text": "Across 155 small-model experiments centered on Qwen 3.5 4B, the same thing kept working: give the model something executable it can check against the evidence it has, and it punches far above its benchmark weight. Here is the field guide — the levers that worked, how I know they're real, and the frontier they opened up.", "date_published": "2026-06-28T00:00:00.000Z", "date_modified": "2026-06-28T00:00:00.000Z", "tags": [ "Essays", "reproducibility", "evaluation", "fine-tuning", "methodology", "llm", "agents" ] }, { "id": "https://frontiercheckpoint.com/essays/agent-harness-evals-moat/", "url": "https://frontiercheckpoint.com/essays/agent-harness-evals-moat/", "title": "The Harness Is the Product: Why Agent Evals Are the Real Moat", "summary": "Swapping the frontier model rarely moves your agent's success rate as much as fixing retries and context management — and the one thing competitors can't clone is your evaluation environment. A thesis on why agent evals, not weights, are where reproducible capability accrues.", "content_text": "Swapping the frontier model rarely moves your agent's success rate as much as fixing retries and context management — and the one thing competitors can't clone is your evaluation environment. A thesis on why agent evals, not weights, are where reproducible capability accrues.", "date_published": "2026-06-27T00:00:00.000Z", "date_modified": "2026-06-27T00:00:00.000Z", "tags": [ "Essays", "agents", "agent-harness", "tool-use", "evaluation" ] }, { "id": "https://frontiercheckpoint.com/signals/vllm-paged-attention-serving-standard/", "url": "https://frontiercheckpoint.com/signals/vllm-paged-attention-serving-standard/", "title": "vLLM and the new default shape of LLM serving", "summary": "If you are still serving with naive static batching, the gap is not marginal — paged KV-cache and continuous batching change the throughput-per-GPU math, and most other stacks have copied the idea.", "content_text": "If you are still serving with naive static batching, the gap is not marginal — paged KV-cache and continuous batching change the throughput-per-GPU math, and most other stacks have copied the idea.", "date_published": "2026-06-26T00:00:00.000Z", "date_modified": "2026-06-26T00:00:00.000Z", "tags": [ "Signals", "serving", "paged-attention", "kv-cache" ] }, { "id": "https://frontiercheckpoint.com/signals/flashattention-3-hopper/", "url": "https://frontiercheckpoint.com/signals/flashattention-3-hopper/", "title": "FlashAttention-3: async, low-precision, Hopper-native", "summary": "The headline is hardware-specific: FA3 is a Hopper story (async copy/MMA overlap, FP8 paths). The portable lesson from the FA line is still the one that matters — attention is bandwidth-bound, and the win is in HBM traffic, not FLOPs.", "content_text": "The headline is hardware-specific: FA3 is a Hopper story (async copy/MMA overlap, FP8 paths). The portable lesson from the FA line is still the one that matters — attention is bandwidth-bound, and the win is in HBM traffic, not FLOPs.", "date_published": "2026-06-24T00:00:00.000Z", "date_modified": "2026-06-24T00:00:00.000Z", "tags": [ "Signals", "flash-attention", "kernels", "gpu-memory" ] }, { "id": "https://frontiercheckpoint.com/essays/economics-of-test-time-compute/", "url": "https://frontiercheckpoint.com/essays/economics-of-test-time-compute/", "title": "The Economics of Thinking: Test-Time Compute as a Scaling Axis", "summary": "Reasoning models turned inference into a per-request dial. This is an economic read on when spending FLOPs at test time actually buys accuracy, why it only pays where answers are cheap to verify, and what variable-cost inference does to latency budgets and capacity planning.", "content_text": "Reasoning models turned inference into a per-request dial. This is an economic read on when spending FLOPs at test time actually buys accuracy, why it only pays where answers are cheap to verify, and what variable-cost inference does to latency budgets and capacity planning.", "date_published": "2026-06-23T00:00:00.000Z", "date_modified": "2026-06-23T00:00:00.000Z", "tags": [ "Essays", "test-time-compute", "reasoning", "serving", "industry" ] }, { "id": "https://frontiercheckpoint.com/signals/mamba-ssm-attention-alternative/", "url": "https://frontiercheckpoint.com/signals/mamba-ssm-attention-alternative/", "title": "Mamba and the selective-state-space line", "summary": "Worth understanding even if you ship transformers: SSMs change the asymptotics (linear in sequence length, constant state at inference) and the failure modes. The interesting deployments are hybrids, not pure-SSM.", "content_text": "Worth understanding even if you ship transformers: SSMs change the asymptotics (linear in sequence length, constant state at inference) and the failure modes. The interesting deployments are hybrids, not pure-SSM.", "date_published": "2026-06-20T00:00:00.000Z", "date_modified": "2026-06-20T00:00:00.000Z", "tags": [ "Signals", "state-space-models", "attention", "long-context" ] }, { "id": "https://frontiercheckpoint.com/explainers/reading-a-model-release/", "url": "https://frontiercheckpoint.com/explainers/reading-a-model-release/", "title": "Reading a Model Release Like an Engineer: Weights, Licenses, System Cards, and Evals", "summary": "The headline benchmark is the least durable thing in a model release. Here is how to read access, licenses, cards, eval protocols, and serving facts before you commit engineering to a number you cannot reproduce.", "content_text": "The headline benchmark is the least durable thing in a model release. Here is how to read access, licenses, cards, eval protocols, and serving facts before you commit engineering to a number you cannot reproduce.", "date_published": "2026-06-19T00:00:00.000Z", "date_modified": "2026-06-19T00:00:00.000Z", "tags": [ "Explainers", "evaluation", "reproducibility", "llm", "industry" ] }, { "id": "https://frontiercheckpoint.com/signals/sglang-radixattention-prefix-cache/", "url": "https://frontiercheckpoint.com/signals/sglang-radixattention-prefix-cache/", "title": "SGLang and RadixAttention for prefix reuse", "summary": "If your workload has heavy shared prefixes — system prompts, few-shot exemplars, agent scaffolds — automatic prefix caching is close to free latency. This is where serving for agents diverges from serving for chat.", "content_text": "If your workload has heavy shared prefixes — system prompts, few-shot exemplars, agent scaffolds — automatic prefix caching is close to free latency. This is where serving for agents diverges from serving for chat.", "date_published": "2026-06-17T00:00:00.000Z", "date_modified": "2026-06-17T00:00:00.000Z", "tags": [ "Signals", "serving", "kv-cache", "agents" ] }, { "id": "https://frontiercheckpoint.com/reproductions/reproducing-nanogpt-speedrun/", "url": "https://frontiercheckpoint.com/reproductions/reproducing-nanogpt-speedrun/", "title": "Reproducing the nanoGPT Speedrun: What Actually Moves the Loss Curve", "summary": "The nanoGPT speedrun is a rare, fully open optimization target: hit 3.28 FineWeb validation loss on a GPT-2 (124M)-class model in minimum wall-clock on 8×H100. We reproduce the pipeline, isolate what the Muon optimizer and the architecture changes actually buy, and flag what will not transfer off the bench.", "content_text": "The nanoGPT speedrun is a rare, fully open optimization target: hit 3.28 FineWeb validation loss on a GPT-2 (124M)-class model in minimum wall-clock on 8×H100. We reproduce the pipeline, isolate what the Muon optimizer and the architecture changes actually buy, and flag what will not transfer off the bench.", "date_published": "2026-06-15T00:00:00.000Z", "date_modified": "2026-06-15T00:00:00.000Z", "tags": [ "Reproductions", "reproducibility", "pretraining", "optimization", "distributed-training" ] }, { "id": "https://frontiercheckpoint.com/signals/deepseek-r1-rl-reasoning-open-weights/", "url": "https://frontiercheckpoint.com/signals/deepseek-r1-rl-reasoning-open-weights/", "title": "DeepSeek-R1: RL-trained reasoning with open weights", "summary": "The reproducible part is the method, not a leaderboard cell: group-relative RL on verifiable rewards, with open weights to probe. It is the cleanest public artifact for understanding the reasoning-model training loop.", "content_text": "The reproducible part is the method, not a leaderboard cell: group-relative RL on verifiable rewards, with open weights to probe. It is the cleanest public artifact for understanding the reasoning-model training loop.", "date_published": "2026-06-13T00:00:00.000Z", "date_modified": "2026-06-13T00:00:00.000Z", "tags": [ "Signals", "reasoning", "grpo", "rlhf" ] }, { "id": "https://frontiercheckpoint.com/signals/modded-nanogpt-muon-speedrun/", "url": "https://frontiercheckpoint.com/signals/modded-nanogpt-muon-speedrun/", "title": "The modded-nanogpt speedrun and the Muon optimizer", "summary": "A rare fully-public optimization target with a reproducible harness — exactly the kind of artifact we like. The Muon optimizer it popularized is the most interesting practical idea to come out of it.", "content_text": "A rare fully-public optimization target with a reproducible harness — exactly the kind of artifact we like. The Muon optimizer it popularized is the most interesting practical idea to come out of it.", "date_published": "2026-06-11T00:00:00.000Z", "date_modified": "2026-06-11T00:00:00.000Z", "tags": [ "Signals", "pretraining", "optimization", "reproducibility" ] }, { "id": "https://frontiercheckpoint.com/libraries/trl-sft-dpo-grpo-library/", "url": "https://frontiercheckpoint.com/libraries/trl-sft-dpo-grpo-library/", "title": "TRL in Anger: SFT, DPO, and GRPO Without Rewriting Your Training Loop", "summary": "TRL turns SFT, DPO, and GRPO into Trainer subclasses that inherit the entire Hugging Face stack — accelerate, peft, DeepSpeed. The convenience is real; the cost is that you're debugging someone else's training loop the moment your problem stops looking like the quickstart.", "content_text": "TRL turns SFT, DPO, and GRPO into Trainer subclasses that inherit the entire Hugging Face stack — accelerate, peft, DeepSpeed. The convenience is real; the cost is that you're debugging someone else's training loop the moment your problem stops looking like the quickstart.", "date_published": "2026-06-11T00:00:00.000Z", "date_modified": "2026-06-11T00:00:00.000Z", "tags": [ "Libraries", "dpo", "grpo", "peft", "fine-tuning", "rlhf" ] }, { "id": "https://frontiercheckpoint.com/explainers/post-training-quantization-gptq-awq-fp8/", "url": "https://frontiercheckpoint.com/explainers/post-training-quantization-gptq-awq-fp8/", "title": "Post-Training Quantization in Practice: GPTQ, AWQ, and FP8", "summary": "Post-training quantization is the cheapest inference lever and the easiest to pull wrong. The right method is set by your serving regime — bandwidth-bound decode wants weight-only INT4, compute-bound prefill wants FP8 — and the win is real only if a fast kernel accelerates your exact config.", "content_text": "Post-training quantization is the cheapest inference lever and the easiest to pull wrong. The right method is set by your serving regime — bandwidth-bound decode wants weight-only INT4, compute-bound prefill wants FP8 — and the win is real only if a fast kernel accelerates your exact config.", "date_published": "2026-06-08T00:00:00.000Z", "date_modified": "2026-06-08T00:00:00.000Z", "tags": [ "Explainers", "quantization", "serving", "kv-cache" ] }, { "id": "https://frontiercheckpoint.com/explainers/grpo-group-relative-policy-optimization/", "url": "https://frontiercheckpoint.com/explainers/grpo-group-relative-policy-optimization/", "title": "GRPO, Demystified: Group-Relative Policy Optimization for Reasoning Models", "summary": "GRPO swaps PPO's learned critic for a Monte-Carlo baseline — the mean reward over a group of sampled completions — trading rollout compute and per-token credit assignment for a simpler, more stable RL loop on verifiable-reward tasks.", "content_text": "GRPO swaps PPO's learned critic for a Monte-Carlo baseline — the mean reward over a group of sampled completions — trading rollout compute and per-token credit assignment for a simpler, more stable RL loop on verifiable-reward tasks.", "date_published": "2026-06-04T00:00:00.000Z", "date_modified": "2026-06-04T00:00:00.000Z", "tags": [ "Explainers", "grpo", "rlhf", "ppo", "reasoning", "fine-tuning" ] }, { "id": "https://frontiercheckpoint.com/explainers/moe-routing-practitioners-guide/", "url": "https://frontiercheckpoint.com/explainers/moe-routing-practitioners-guide/", "title": "Routing Is the Hard Part: A Practitioner's Guide to Mixture-of-Experts", "summary": "MoE decouples parameter count from per-token FLOPs, but every hard problem — instability, dropped tokens, load imbalance, all-to-all traffic, a footprint set by total not active params — lives in the router. A structural tour from Switch/GShard to fine-grained and aux-loss-free designs, and the systems bill you actually pay.", "content_text": "MoE decouples parameter count from per-token FLOPs, but every hard problem — instability, dropped tokens, load imbalance, all-to-all traffic, a footprint set by total not active params — lives in the router. A structural tour from Switch/GShard to fine-grained and aux-loss-free designs, and the systems bill you actually pay.", "date_published": "2026-06-01T00:00:00.000Z", "date_modified": "2026-06-01T00:00:00.000Z", "tags": [ "Explainers", "mixture-of-experts", "transformers", "tensor-parallelism" ] }, { "id": "https://frontiercheckpoint.com/recreations/recreating-flashattention-tiled-kernel/", "url": "https://frontiercheckpoint.com/recreations/recreating-flashattention-tiled-kernel/", "title": "Recreating FlashAttention: A Tiled, IO-Aware Attention Kernel from Scratch", "summary": "FlashAttention is exact attention restructured for the memory hierarchy, not an approximation. We implement the tiled forward and recompute backward in Triton, validate exactness against a reference, and separate what a tutorial actually reproduces from what needs CUTLASS-grade engineering.", "content_text": "FlashAttention is exact attention restructured for the memory hierarchy, not an approximation. We implement the tiled forward and recompute backward in Triton, validate exactness against a reference, and separate what a tutorial actually reproduces from what needs CUTLASS-grade engineering.", "date_published": "2026-05-28T00:00:00.000Z", "date_modified": "2026-05-28T00:00:00.000Z", "tags": [ "Recreations", "flash-attention", "kernels", "attention", "gpu-memory" ] }, { "id": "https://frontiercheckpoint.com/explainers/rope-long-context-stack/", "url": "https://frontiercheckpoint.com/explainers/rope-long-context-stack/", "title": "RoPE and the Long-Context Stack: Rotation, Interpolation, and What Breaks at 128k", "summary": "RoPE turns position into a per-dimension rotation — and that same rotation is why PI, NTK-aware scaling, and YaRN exist, and why a 128k window rarely means 128k of usable context. The math, the methods, and the serving bill.", "content_text": "RoPE turns position into a per-dimension rotation — and that same rotation is why PI, NTK-aware scaling, and YaRN exist, and why a 128k window rarely means 128k of usable context. The math, the methods, and the serving bill.", "date_published": "2026-05-24T00:00:00.000Z", "date_modified": "2026-05-24T00:00:00.000Z", "tags": [ "Explainers", "rope", "long-context", "attention", "transformers" ] }, { "id": "https://frontiercheckpoint.com/libraries/vllm-paged-attention-serving/", "url": "https://frontiercheckpoint.com/libraries/vllm-paged-attention-serving/", "title": "vLLM, Explained: PagedAttention, Continuous Batching, and the Serving Stack", "summary": "vLLM treats the KV cache like OS virtual memory — non-contiguous paged blocks — and schedules work at the token, not the request. You get high aggregate throughput; the cost is that per-request latency becomes something you tune rather than something you get for free.", "content_text": "vLLM treats the KV cache like OS virtual memory — non-contiguous paged blocks — and schedules work at the token, not the request. You get high aggregate throughput; the cost is that per-request latency becomes something you tune rather than something you get for free.", "date_published": "2026-05-20T00:00:00.000Z", "date_modified": "2026-05-20T00:00:00.000Z", "tags": [ "Libraries", "paged-attention", "kv-cache", "serving", "llm" ] }, { "id": "https://frontiercheckpoint.com/explainers/distributed-training-fsdp-zero-parallelism/", "url": "https://frontiercheckpoint.com/explainers/distributed-training-fsdp-zero-parallelism/", "title": "Sharding the Model: FSDP, ZeRO, and Tensor/Pipeline Parallelism", "summary": "Past one GPU you stop training a model and start operating a distributed system. Here is what each parallelism axis actually shards, what it costs on the wire, and how practitioners stack them into 3D/4D layouts.", "content_text": "Past one GPU you stop training a model and start operating a distributed system. Here is what each parallelism axis actually shards, what it costs on the wire, and how practitioners stack them into 3D/4D layouts.", "date_published": "2026-05-15T00:00:00.000Z", "date_modified": "2026-05-15T00:00:00.000Z", "tags": [ "Explainers", "distributed-training", "fsdp", "tensor-parallelism", "gpu-memory" ] }, { "id": "https://frontiercheckpoint.com/essays/the-checkpoint-signal-vs-noise/", "url": "https://frontiercheckpoint.com/essays/the-checkpoint-signal-vs-noise/", "title": "How We Separate Signal From Noise: Frontier Checkpoint's Verification Rubric", "summary": "The standard behind everything we publish: the filters that decide what earns your attention, the reproduced-to-unverified ladder we grade claims on, how we handle benchmarks and weightless releases, and why every correction is dated and logged rather than silently edited.", "content_text": "The standard behind everything we publish: the filters that decide what earns your attention, the reproduced-to-unverified ladder we grade claims on, how we handle benchmarks and weightless releases, and why every correction is dated and logged rather than silently edited.", "date_published": "2026-05-12T00:00:00.000Z", "date_modified": "2026-05-12T00:00:00.000Z", "tags": [ "Essays", "methodology", "reproducibility", "evaluation", "industry" ] } ] }