AI model
Dense 27-billion-parameter language model in Alibaba's Qwen3.8 family, small enough to run on local hardware and used as a base for fine-tunes.
BAAI's AREX-2 config declares one multi-token-prediction layer, but its weight index lists 0 such tensors against Qwen3.8-27B's 15. Its dev.to testers say the base-model drafter and quant need rechecking too, though they have not run AREX-2.
Reality
EcoHash reports that one vLLM flag took Qwen3.8-27B on a single RTX Pro 6000 from 999.9 to 1954.0 output tokens/sec. The figure comes from its own production config on vLLM 0.27.1, so it is one team's measurement on its own workloads.
Reality
Sigil Wen's Underdog runs a 27-billion-parameter AI assistant entirely on users' own Macs and Windows PCs. Privacy then depends on where the model runs, and users pay for it with a smaller model and a vendor that earns when the assistant spends.
Reality
Strata runs a 125B-class Qwen model at about 100 tok/s on one RTX 4090 because its sparse MoE activates only about 6B parameters per token. The design moves the hardware bill to system RAM, and the 100 tok/s figure holds mainly for code and structured output.
Reality
Jeff, an open 0.8B model, returns a probability for each option in one forward pass and sends low-confidence agent decisions to Qwen3.8-27B. The confidence score attached to each answer tells an agent when a small decision is worth the larger model's time.
Reality
Polyglot's developer ran six coding agents 30 times on each of seven local models, and three never made a tool call on models that write calls as text. The author's own error bars say 30 runs can sort agents into tiers but cannot rank two agents inside one.
Reality
Researchers writing on LessWrong found only Nemotron 3 Super and Qwen3 32B changed refusals when test suspicion was removed from their reasoning traces. In most other models, test talk in a trace looks like general caution, so it is weak evidence of eval gaming.
Reality
Portable Computer's harness is real engineering. The bill of materials starts at a $4,800 DGX Spark or a 24GB RTX card, which turns local agents into a hardware order.
Perspective Coverage
3 publishersReality
LM Studio 0.4.0 shipped llmster, a headless daemon that runs it on the Linux GPU servers where MIT-licensed Ollama already worked. Teams whose policy demands auditable source now decide on licence, since only LM Studio's lms CLI carries an MIT grant.
Reality
The startup says Anchor 3.0 caught more than 90 percent of violations in a test it published itself, at under a 500th of the cost of a frontier call, and the messages it misses stay the firm's problem.
Reality
Alibaba's Qwen3.8-Flash-Next preview activates 6B of its 125B parameters per token. Per-token compute drops to under a quarter of the dense 27B's, and about 125GB of weights still has to stay on device.
Reality
Bonsai 2 27B keeps 98 percent of Qwen3.8 27B's aggregate benchmark score, up from 95 percent for PrismML's first release, and TechCrunch reports the 5.9GB file is small enough for a PC and possibly a high-end phone.
Reality
The merged 27B finished 37 of 100 internal coding tasks against 34 for the fast checkpoint and 39 for the reasoning one, and it did that while emitting 71% fewer output tokens than the 39, with no post-training in the recipe.
Reality
PrismML's ternary build of Qwen3.8 27B keeps 98.2 percent of the full-precision benchmark average on both of the company's inconsistent scorecards, and the loss it does take is concentrated in knowledge and reasoning.
Reality
Atlassian's own validator did the grading, on 25 app briefs, on the same workstation that trained the adapter. Anyone reusing the pass rate needs briefs that look like those 25.
Reality
The calculator prices its featured coding workload at about $6.94 a month through OpenRouter against 26 cents of electricity. The machine is still $3,419 down after a year and 7.97 billion tokens from break-even.
Reality
Every task in the benchmark is checked against its required terminal state, so any tool order can pass and any extra side effect fails. Claude Opus 5, the strongest model tested, cleared 66.50% at pass@1.
Reality
A 15M-parameter model streams English text on a 2007 PSP at about one token per second. That is the extreme end of a sizing rule. The harder half of that rule is checking whether the file that fits is a format its own maintainer recommends.
Reality
Nemotron 3.5 Lightning touches a tenth of its weights per token, which is what makes an agent loop plausible next to the sensors, but memory still has to hold all thirty billion and NVIDIA's own loop still escalates off-device.
Reality
Prefill and decode sit on the same card and answer to different limits, which is why a GPU with more arithmetic and the same memory read rate leaves your time per output token exactly where it was.
Reality
Build · August 24, 2026 · Confirmed2 publishers
Build · August 19, 2026 · Confirmed2 publishers