Releases·Tool Release
Improved chat template for Laguna XS / S 2.1 (configurable forced thinking, preserve_thinking toggle, and stability fixes)
A community member released an updated Jinja chat template for Poolside's Laguna XS and S 2.1 models, motivated by inconsistent reasoning — the models often skip thinking when it would help, especially the XS variant. The template adds force_thinking, reasoning_effort presets, a preserve_thinking toggle, and control-token validation, while defaulting to upstream behavior as a drop-in replacement.
17 engagement1 sourceSun, Sep 27, 2026, 04:10 PM
AI Overview
- Template adds force_thinking toggle, reasoning_effort presets (none/auto/max), and reasoning preservation that doesn't invalidate prefix caches.
- Original template's forced history preservation bloats context; new preserve_thinking toggle can be disabled for normal chats.
- Adds validation to catch smuggled control tokens across roles, disableable via allow_injection: true.
- Defaults match upstream behavior (enable_thinking=true, preserve_thinking=true, force_thinking=false) for drop-in replacement.
- Template repo includes recommended llama.cpp sampling settings and a Pi models. config snippet.
- Poster reports Laguna models skip reasoning inconsistently, especially XS 2.1, but think well when prompted; tags post-train lead u/matthiasgalle.
Related
Social·Community
Community reports chat template bugs breaking reasoning in Laguna-S-2.1 and DeepSeek-V4
Users report that chat template bugs in llama.cpp and Ollama break reasoning behavior in Laguna-S-2.1 and DeepSeek-V4, causing models to skip thinking or emit empty tags. The community has identified fixes, including updated chat templates and GGUFs, which restore proper reasoning and tool calling.
Users report that chat template bugs in llama.cpp and Ollama break reasoning behavior in Laguna-S-2.1 and DeepSeek-V4, causing models to skip thinking or emit empty tags. The community has identified fixes, including updated chat templates and GGUFs, which restore proper reasoning and tool calling.
Updated Fri, Jul 31, 2026, 08:27 PM — Poolside released updated GGUFs and chat template fixing reasoning and tool calling in Laguna-S-2.1.
113290 engagement6 sourcesredditWed, Jul 22, 2026, 07:56 AM
Social·Community
Users report Laguna S 2.1 overthinking and looping at low quantization, fixed by MoE-aware quants
Multiple users report that the Laguna S 2.1 model exhibits excessive thinking loops or overthinking when run at low-bit quantization (e.g., Q2_K_XL, IQ3_S). One user found that switching to an MoE-aware APEX quant (e.g., Myric/Laguna-S-2.1-APEX-GGUF) resolved the issue, while another running Q8 reports no problems. The behavior appears to be a quantization artifact, not a model defect.
Multiple users report that the Laguna S 2.1 model exhibits excessive thinking loops or overthinking when run at low-bit quantization (e.g., Q2_K_XL, IQ3_S). One user found that switching to an MoE-aware APEX quant (e.g., Myric/Laguna-S-2.1-APEX-GGUF) resolved the issue, while another running Q8 reports no problems. The behavior appears to be a quantization artifact, not a model defect.
Updated Tue, Jul 28, 2026, 04:17 PM — A user reports that switching to an MoE-aware APEX quant (Myric/Laguna-S-2.1-APEX-GGUF) resolves the thinking loops, con
130215 engagement4 sourcesredditThu, Jul 23, 2026, 06:52 AM
Releases·Tool Release
First Tree launches OpenTag, open-source Slack AI coworkers running on existing Claude Code, Codex and Pi subscriptions
The developers released OpenTag, an open-source AI coworker for Slack that delegates background tasks to agents running on users' own machines and existing coding-agent subscriptions. It targets frustration with closed-source, cloud-hosted, per-seat AI coworker tools that double-bill inference. The project is pre-alpha, self-hostable, and seeking community feedback.
The developers released OpenTag, an open-source AI coworker for Slack that delegates background tasks to agents running on users' own machines and existing coding-agent subscriptions. It targets frustration with closed-source, cloud-hosted, per-seat AI coworker tools that double-bill inference. The project is pre-alpha, self-hostable, and seeking community feedback.
1 engagement1 sourceredditMon, Sep 14, 2026, 10:43 PM
Social·Community
User benchmarks DFlash speculative decoding on Laguna S 2.1: 2.5x slower default, tuned to match baseline
A user tested DFlash speculative decoding on Laguna S 2.1 (118B MoE, 71 GB Q4) with 2× RTX 5090. Default flags yielded 23 tok/s (2.5× slower than the 58 tok/s baseline without draft). After tuning, they achieved 64 tok/s, slightly above the 62 tok/s baseline. The model's 71 GB exceeds the 64 GB VRAM, causing expert spillover to CPU RAM.
A user tested DFlash speculative decoding on Laguna S 2.1 (118B MoE, 71 GB Q4) with 2× RTX 5090. Default flags yielded 23 tok/s (2.5× slower than the 58 tok/s baseline without draft). After tuning, they achieved 64 tok/s, slightly above the 62 tok/s baseline. The model's 71 GB exceeds the 64 GB VRAM, causing expert spillover to CPU RAM.
17 engagement1 sourceredditThu, Jul 23, 2026, 03:12 PM
Releases·Tool Release
llama.cpp PR adds adaptive MTP depth selection
A llama.cpp contributor filed PR#27210 adding an adaptive multi-token prediction (MTP) mode that dynamically sets MTP depth via a counting-style state machine, removing the need to manually tune depth. Early benchmarks show roughly 3% performance loss versus fixed MTP depth 3 on dense prose, with potential gains in some scenarios.
A llama.cpp contributor filed PR#27210 adding an adaptive multi-token prediction (MTP) mode that dynamically sets MTP depth via a counting-style state machine, removing the need to manually tune depth. Early benchmarks show roughly 3% performance loss versus fixed MTP depth 3 on dense prose, with potential gains in some scenarios.
199 engagement1 sourceredditMon, Aug 17, 2026, 06:05 PM