buildConfirmed3 publishers Ai2 open-sourced AstaBrief 8B, which writes cited research reports in 51.1 seconds against 178.5 for Asta's Claude-powered mode. Labs that cannot send unpublished research questions to a hosted model can now run a cited-report generator on their own servers.
Perspective Coverage
3 publishers
- Builder
- Builder 55%
- Operator
- Operator 33%
- Investor
- Investor 12%
Reality
- Evidence55
- Adoption18
- Hype gap+22
- Incentives55
- Confidence60
buildOne report1 publisher SageMaker's fallback instance can serve a quantized copy of the same model
AWS counted 13 SageMaker inference launches so far in 2026. The one that changes production behaviour most is a prioritized list of up to five instance types, each allowed its own model optimization settings.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives85
- Confidence55
An 11-model probe puts the "this is impossible" signal 85 degrees from the refusal direction
Yucheng Du and Xiyang Hu report a single direction that separates answerable from unanswerable math and code prompts at 0.939 mean AUC across 11 models, and its mean cosine with the canonical safety-refusal direction is 0.087.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+12
- Incentives60
- Confidence50
buildOne report1 publisher ShadowPEFT reaches PEFT's main branch as an adapter that keeps its own hidden state
ShadowPEFT is in Hugging Face PEFT's main branch as of a September 15th announcement, and it carries its own hidden state and can be detached as a smaller standalone model. Trying it means installing PEFT from source.
Reality
- Evidence58
- Adoption20
- Hype gap+10
- Incentives65
- Confidence57
buildOne report1 publisher Running llama-server puts context, KV cache and GPU placement in your command line
A dev.to guide puts the llama.cpp and Ollama decision on ownership. Under Ollama the named model is the unit you operate; under llama.cpp it is the llama-server process and the seven flags in its command line.
Reality
- Evidence52
- Adoption
- Not scored
- Hype gap0
- Incentives30
- Confidence58
buildOne report1 publisher One Docker log fixture carries the entire margin in a thirty-run local-model benchmark
The coding-trained model came out ahead over thirty runs, but four of the five fixtures tied, which leaves the whole margin in a single Docker log task run at default temperature between a 30b model and an 8b one.
Reality
- Evidence58
- Adoption10
- Hype gap−15
- Incentives25
- Confidence55
buildOne report1 publisher Escalating retries recover the 7,930 refusal prompts a single steering pass drops
Multiverse Computing's paper treats a deployment's refusal set as a subset of politics rather than the whole topic, which changes what the training corpus has to contain before any model is trained. The posted text breaks off before the results.
Reality
- Evidence42
- Adoption12
- Hype gap−20
- Incentives65
- Confidence58
Fine-tuning hands the open-weight safety question to the buyer
The 270-company letter and Anthropic's warning are describing the same lever from opposite ends. Fine-tuning is what lifts a small open model past a frontier one, and it is also what can undo the safeguards shipped with it.
Reality
- Evidence26
- Adoption
- Not scored
- Hype gap+32
- Incentives72
- Confidence28
Reasoning models spend extra tokens when a prompt cuts against a stereotype
A team adapting the implicit association test to reasoning traces found four of five models working harder on association-incompatible prompts, which puts a measurable bias signal in the process rather than only in the answer.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+14
- Incentives45
- Confidence55
buildOne report1 publisher Qwen3 4B matches Qwen2.5 7B case for case at half the cold start
Across 20 paired writing-correction cases on Ollama for Windows, the two models succeeded on the same 18 and failed on the same 2, while the 4B averaged 23.99 seconds of cold start against 54.37. That reorders local shortlisting.
Reality
- Evidence46
- Adoption14
- Hype gap+16
- Incentives42
- Confidence55
buildOne report1 publisher Four months of A100 bills say self-hosting is a utilization bet, not a cost saving
One engineer's home lab tally: $1,400 a month for two A100s running 40% idle, against open-weight models he measured inside noise of GPT-4o. The break-even is real, and it sits high.
Reality
- Evidence22
- Adoption10
- Hype gap+36
- Incentives66
- Confidence34