Quinn 3.8 27B decoded at 39.1 tokens a second on an M1 Max in a developer’s test, up from 18.9 across five prompts. That is a striking jump. The comparison starts with stock Splash forced onto a chip it does not officially support: Splash requires an M3 or newer.
The M1 Port Improved More Than Decode
The community port changed more than a chip check. It added a decode kernel using 32-bit floats and a prefill tile that uses the M1’s matrix-multiply units. The developer measured gains in both decode and prefill.
For a 2048-token prompt, prefill rose from 53 to 142 tokens a second. With four requests running at once, total speed rose from 18.0 to 65.2 tokens a second. The headline decode number is only one part of what the developer tested.
The output was close to the previous M1 build, but not identical. Across 2528 positions, the model’s top next-token choice matched 99.92% of the time. The release also supports GGUF and splits long prompts into chunks.
Memory puts another limit on who can use the full model. Inco lists 36 gigs or more for the full 27B package, with 48 recommended. The port’s 16-gig figure is for a much smaller model. Taken together, these are large developer-measured gains on an M1 Max, with an unsupported stock Splash run as the starting point.
DeepSeek Measured Operations, Not a Whole Model
DeepSeek released Deep Gem Ascend with familiar function calls and code for Huawei Ascend 950. Its supported matrix operations include BF16, FP8, and FP4. The chip-specific work moves and arranges numbers, then overlaps loading with calculation where possible.
Those familiar entry points apply to the operations the code supports. Masked grouped matrix multiplication interfaces are marked unimplemented, and surrounding integrations may need work. The release does not guarantee that an entire CUDA application will move unchanged.
DeepSeek reports strong results for selected matrix multiplications. These are its own measurements of individual operations. They do not show how fast a whole model runs, or establish a matched win against Nvidia. A fast operation is useful to know about, but it answers a narrower question than whole-model speed.
The Coding Harness Changed the Score
Coding-agent results have a similar boundary. On 63 terminal tasks, Claude Opus 5 scored 49% in Claude Code and 30% in PI. GPT-6 Astra scored 56% in Codex and 60% in PI. Each model got different scores across the harnesses used in this task set.
The comparison does not isolate which part of a harness caused the difference. It does show why the model name alone does not describe the whole test. The tools around an agent’s answer are part of what a reader needs to understand.
Several updates focus on that surrounding work. Claude Code added a built-in “You should know” plugin that scans agent output for a detail a reader might miss. Codex CLI 0.160.0 added an opt-in Guardian review that can use earlier instructions from agent handoffs. Cursor Rollouts can identify an offending PR, open an issue, and start a cloud agent to fix it.
Those features address different points in a workflow: noticing a detail in an answer, carrying handoff instructions into review, and beginning a fix after a PR is identified. The 63-task comparison gives a concrete reason to pay attention to the harness, while leaving open which of its parts affected those scores.
Read the Agent Demonstration as Shown
Microsoft Autopilot remains in expanded private preview. Microsoft reported more than 450 million paid Microsoft 365 commercial seats in January, but @NateBJones points out that those seats are not all paying for AI. In Microsoft’s customer renewal demonstration, a person begins by asking a question. The demonstration does not show Autopilot finding the problem on its own.
@NateBJones distinguishes a repeated briefing from an agent responsible for following an assignment between check-ins. For the renewal example, he says to check the links behind important claims and whether the tool could access the relevant email, Teams conversation, and meeting notes. If the answer is weak, missing information or permissions could be part of the reason, alongside how the model handled what it saw.
Other accounts describe finished work, with boundaries of their own. @NateBJones says Opus 5.5 helped produce a 500-piece Lego build, a video with sound, a 63-page instruction booklet, and assembly stills. He says he checked every piece and connection. The job used 89 million tokens; he reports 50 cents in actual subscription costs, versus 40 or 50 dollars in estimated API costs.
Anthropic designer Nate showed a shorter path from an idea to an app: ask Claude Design for onboarding options, edit the details himself, try a click-through prototype, then ask Claude to build it. Carter Church reports that GPT6 Astra helped recover a letter dated to March 1809. He published the key, transcription, and reproduction scripts; Tmokio marked the solution solved after review, though five signs remain uncertain. Church reran the solver without Napoleonic text in its scoring material. That check does not establish what Astra encountered during training.