Open Weights Split

If You Only Read One Thing

Open-weight AI is splitting into two useful shapes, not marching toward one frontier clone. MiniMax-M3 Gets Judged shows the 1M-context, agentic frontier path; Liquid Shrinks the Agent shows the local tool-caller path. The useful read is Artificial Analysis on MiniMax-M3, because the abstention numbers turn "open weights" from a release label into a routing problem for every agent stack.

MiniMax-M3 Gets Judged

MiniMax-M3 was already important when it launched. The new fact is that independent measurement makes it look less like "open weights catch the frontier" and more like "open weights are learning to pick their battles."

Last week's AI briefing treated M3 as a long-context architecture story: MiniMax promised a 1M-token context window, native image and video input, desktop operation, and open weights to follow after the API launch. The official M3 page still frames the model around coding, agentic benchmarks, and MiniMax Sparse Attention, the sparse-attention mechanism meant to make million-token sessions economically usable. The June 8 addition is Artificial Analysis' evaluation: M3 scores 55 on its Intelligence Index, ahead of open-weight peers Kimi K2.6 and MiMo-V2.5-Pro at 54, but behind GPT-5.5 and Claude Opus 4.8 on the real-work GDPval-AA slice.

Why it matters: The most interesting number is not the headline score. It is the refusal shape. Artificial Analysis says M3 attempts only 30.9% of AA-Omniscience questions, the lowest among current peers, which gives it a 16.1% hallucination rate but only 15.0% accuracy. That is an architectural and product signal: a lower-cost, long-context agent model can look safer by abstaining more aggressively, but the deployment question becomes whether the harness can tell the difference between useful caution and missing work.

The mechanism is evaluation-aware routing. M3's price is $0.30 per million input tokens and $1.20 per million output tokens up to 512K context, then doubles above that. That makes it attractive for long repository reads, planning passes, and retrieval-heavy agent loops where closed frontier models are too expensive to run repeatedly. But heavy abstention means M3 may be better as a planner, reviewer, or long-context filter than as the final authority. The router has to know when cheap context is enough and when to escalate.

Room for disagreement: MiniMax still has not completed the open-weight proof until the weights and technical report are actually out. The API can be evaluated, but the community cannot yet inspect, reproduce, quantize, or self-host the exact artifact. If the weights land with a restrictive license or worse reproduction quality, the "open" part becomes distribution marketing, not deployment freedom.

Liquid Shrinks the Agent

Liquid AI is making the opposite bet from MiniMax. Instead of stretching the frontier model to a million tokens, it is trying to make the local agent fast enough that the cloud call becomes optional.

The LFM2.5-8B-A1B release is an 8.3B-parameter mixture-of-experts model with roughly 1.5B active parameters per token, 128K context, and open weights available through Hugging Face. Liquid says it moved from 12T to 38T training tokens, doubled the tokenizer vocabulary, added large-scale reinforcement learning, and made the post-trained model "reasoning-only," meaning it generates reasoning traces before final answers. The practical hook is day-one support across llama.cpp, MLX, vLLM, SGLang, ONNX, and GGUF.

Why it matters: Small local models have usually failed at the exact thing agents need: reliable tool use across a long menu. Liquid's numbers are aimed at that gap. It reports BFCLv3 rising from 45.07 to 64.36, BFCLv4 from 25.52 to 48.50, and Tau2-Telecom from 13.60 to 88.07 versus the prior LFM2-8B-A1B. It also claims 253 tokens per second on an M5 Max, about 30 tokens per second on a phone, and under 6 GB memory use. Those are vendor numbers, but they are the right vendor numbers: tool calling, latency, memory, and local runtime support.

The mechanism is not "small models beat frontier models." They do not. The mechanism is that agents are not one model call. They are search, classify, select, call, verify, and summarize loops. If a 1.5B-active model can handle the cheap parts locally, then the expensive frontier call becomes a scarce escalation path instead of the default executor. That changes cost, privacy, and interaction latency at the same time.

Room for disagreement: Liquid's own knowledge numbers still show a limited model. Its AA-Omniscience Index is negative, and accuracy remains low even after a large non-hallucination gain. The bet only works if the harness treats LFM2.5 as a local worker with a narrow job, not as a general assistant pretending to know more than it does.

The Contrarian Take

Everyone says: Open-weight progress means the open ecosystem is catching the closed frontier.

Here's why that's wrong (or at least incomplete): The stronger reading is specialization. MiniMax-M3 is a long-context, multimodal, coding-agent candidate with a cost profile that begs for router-based escalation. Liquid's LFM2.5 is a local worker candidate that makes tool dispatch and private loops cheap. Neither story says a generic open model replaces Claude, GPT, or Gemini tomorrow. Both say the frontier call is becoming one tier in a system, not the whole system.

Under the Radar

Claude Code added a recovery boot path - The 2.1.169 changelog adds --safe-mode to start with customizations, skills, hooks, plugins, MCP servers, and project instructions disabled. That is not glamorous, but it is a real reliability primitive: when agent behavior is corrupted by the environment, the runtime now has a clean diagnostic mode.
Factory is selling the router, not the model - Factory Router claims 20-25% lower token spend while preserving 96-99% of Claude Opus 4.7 pass rate on enterprise engineering benchmarks. The important part is the product shape: model selection, provider failover, reserved throughput, and admin routing guidance are being bundled as the coding-agent control plane.

Quick Takes

MiniMax's "open" clock is now the story. MiniMax said the M3 weights and technical report would follow within roughly 10 days of the June 1 launch. After Artificial Analysis' eval, the next test is not another benchmark chart; it is whether self-hosted M3 behaves like the API model. (Source)
Cursor's Design Mode keeps pushing UI feedback into the agent loop. Cursor's June 5 changelog lets users multi-select browser elements and queue voice feedback while an agent is running. That is not another model picker. It is a richer observation channel for frontend agents, where pointing at the screen can replace a brittle written prompt. (Source)
Codex Sites moves generated apps into the work surface. OpenAI's Codex expansion includes Sites and annotations for turning plain-text requests into shareable internal tools and refining outputs against document or site regions. The technical signal is that app generation is becoming a hosted workflow surface, not only a local coding task. (Source)

The Thread

The thread is model unbundling. MiniMax is trying to own the cheap long-context tier. Liquid is trying to own the local worker tier. Factory is routing between tiers. Claude Code is adding recovery state for when the tiered environment goes wrong. Cursor and Codex are turning richer human feedback into agent context. The frontier model still matters, but the daily engineering problem is deciding which part of the loop deserves it, and which part should be handled by cheaper context, local dispatch, or recovery tooling before escalation.

Predictions

New predictions:

I predict: By 2026-07-31, MiniMax-M3's released weights will appear in at least one independent long-context coding or agent benchmark, and the self-hosted result will trail the MiniMax-hosted API by at least 5 percentage points on one major task. (Confidence: medium; Check by: 2026-07-31)
I predict: By 2026-08-31, at least one mainstream coding-agent runtime or IDE agent will ship an explicit local-worker routing mode that uses an 8B-or-smaller open-weight model for tool selection, classification, or draft review before escalating to a frontier model. (Confidence: medium; Check by: 2026-08-31)

Generated: 2026-06-09 03:53 EDT