RAMGPT research paths
Independent Research.
Start with a question.
Follow the evidence, not just the headline.
Choose a topic below to find selected reports and the concepts behind them. These are reading paths, not rankings or an exhaustive list of everything we publish.
Read our research method ↓Reading path
Local inference
Where does the time go on real hardware?
Start with measurements of long context and GPU offloading. Check the hardware, memory budget, and workload before applying a result to your own setup.
Independent measurement
Qwen3.8-27B at 200K: The Performance Cliff Is VRAM, Not IQ3
Independent RTX 4090 testing finds UD-IQ3_XXS only ~2% slower than UD-Q3_K_XL at matched context, while reducing the 200K GPU memory budget from 16 to 15 GiB cuts prompt processing by 19%.
Independent measurement
Adding a Second GPU Made My 4090 Slower on Qwen3.8
An independent Qwen3.8 Flash-Next benchmark on RTX 4090 + RTX 3060 showing how auto-fit, manual MoE placement, quantization, and MTP changed decode speed.
Reading path
Quantization
What changes when a model uses fewer bits?
Compare fidelity, task behavior, and speed without treating them as the same measurement. Each report keeps its own baseline, conditions, and limitations.
Independent reproduction
GSQ-RCO Beats Unsloth on WikiText-2 at ~8.4GB — But Not on Throughput
An RTX 4090 reproduction of Qwen3.8-27B at ~2.5 bpw finds 1.93% lower WikiText-2 perplexity for GSQ-RCO than Unsloth UD-IQ2_S, while tensor dumps reveal a far more heterogeneous precision allocation and slightly lower throughput.
Independent measurement
MiMo-V2.6 QuantBench: Public Q4 and IQ4_XS Compared
We tested public MiMo-V2.6-Distill-Qwen-9B GGUFs on one RTX 4090, comparing loadability, perplexity, speed, exact-task accuracy, and BF16 fidelity.
Reading path
Agents & tool use
How much work happens around the model?
Examine tool-call structure, context overhead, and task scaling. Follow both completion time and output validation, rather than speed alone.
Independent measurement
Gemma 4 vs Qwen3.8: Parallel Tool Calling Changes Agent Scaling
A local RTX 4090 agent benchmark shows Gemma 4 winning small tasks on speed while Qwen3.8 keeps agent turns flat by batching parallel tool calls.
Independent measurement
How to Reduce AI Agent Context and Tool-Call Overhead
A real AI-agent workflow audit shows how task routing, compact state checks, shared validation scripts, and smaller handoffs can reduce context and tool-call overhead.
How we investigate
Our research method
RAMGPT treats technical claims as hypotheses to test, not headlines to repeat. We independently reproduce published and community results, isolate the mechanisms behind them, and publish evidence that readers can inspect and challenge.
What we publish
- Reproduction Reports — independent attempts to reproduce published benchmark or performance claims.
- Claim Audits — checks for hidden assumptions, unfair baselines, benchmark leakage, cache effects, token-budget differences, and other confounders.
- Ablation Studies — controlled experiments that remove or change one mechanism at a time to identify what actually causes an observed gain.
- Performance Forensics — low-level investigation of latency, throughput, memory, storage I/O, page faults, utilization, quantization, and runtime behavior.
- Independent Benchmarks — versioned comparisons designed around exact artifacts, pinned references, reproducible workloads, and inspectable evidence.
- Research Notes — new technical observations discovered while reproducing or extending existing work.
Our experimental standard
A useful experiment starts with a falsifiable claim. We document the hardware, model artifact, software stack, settings, prompts or datasets, measurement method, and important limitations. When the explanation behind a result is uncertain, we prefer matched controls and ablations over speculation. When a result does not reproduce, we report that outcome rather than forcing a positive conclusion.
From reproduction to original research
Reproduction is the starting point, not the endpoint. The most valuable questions often appear after the first result: whether a reported gain survives an equal-token comparison, whether an SSD-offload claim holds under cold-cache conditions, whether a quantization result survives task-level evaluation, or whether an optimization still works on consumer hardware. Those unanswered questions become independent RAMGPT experiments.
Open evidence
Where practical, RAMGPT publishes benchmark inputs, summaries, scripts, manifests, provenance records, and other artifacts needed to inspect the result. RAMGPT QuantBench is the first standing research program built around this model.
Research standards and correction rules are documented in the editorial policy.