📋 Models

by anon · 2026-06-22 12:10:08
← all clips

🧑 You

Imagine you have a coding agent, like opencode or pi. You plan on running local only models. Which models do the community consider the best between the Gemma 4 moe or dense or the Qwen 3.6 moe or dense. Assume I have 32 gig vram and I care deeply about t/s perf. I need the fastest thing that can reliably call tools.

▶ Tool: read_file [path='~/Personal/skills/skill_index.md']

▶ Tool: web_search [max_results={}, query='Gemma 4 MoE vs dense local coding ag]…

▶ Tool: web_search [max_results={}, query='Qwen 3.6 MoE vs dense local model co]…

▶ Tool: web_search [max_results={}, query='Gemma 4 26B MoE vs 31B dense tool ca]…

▶ Tool: web_search [max_results={}, query='Qwen 3.6 35B-A3B MoE vs 27B dense to]…

▶ Tool: fetch_url [url='https://insiderllm.com/guides/best-local-coding-models]…

▶ Tool: fetch_url [url='https://zoliben.com/en/posts/2026-04-23-qwen-36-35b-vs]…

▶ Tool: fetch_url [url='https://zoliben.com/en/posts/2026-04-23-qwen-36-35b-vs]…

▶ Tool: web_search [max_results={}, query='Gemma 4 26B MoE tokens per second RT]…

▶ Tool: fetch_url [url='https://insiderllm.com/guides/wicked-fast-gemma-4-26b-]…

▶ Tool: web_search [max_results={}, query='Gemma 4 26B MoE tool calling vs Qwen]…

▶ Tool: fetch_url [url='https://openclawradar.com/article/qwen3-27b-vs-gemma-4]…

🤖 Assistant

Alright, I dug deep into the community benchmarks, first-hand tests, and r/LocalLLaMA discussions from the past 2 months. Here's the real picture for your 32GB VRAM + OpenCode/Pi + tool-calling + speed-first setup.

---

The Short Answer

For pure speed + reliable tool calling → Qwen 3.6-35B-A3B MoE is your best bet. It's the sweet spot.

If you want maximum reliability at the cost of speed → Qwen 3.6-27B Dense is the community's top pick for agentic coding.

If speed is everything and you can tolerate a tool-call quirk → Gemma 4 26B-A4B MoE is 3x faster than anything else.

---

The Numbers (All measured on real hardware)

Speed (tok/s) — what you said you care about most

Model

RTX 3090 (24GB)

RTX 4090 (24GB)

Your 32GB card (est.)

Gemma 4 26B-A4B MoE (Q4)

128 tok/s

~150 tok/s

~150+ tok/s

Qwen 3.6-35B-A3B MoE (Q4)

~100 tok/s

161 tok/s

~160+ tok/s

Qwen 3.6-27B Dense (Q4)

41 tok/s

40 tok/s

~45 tok/s

Gemma 4 31B Dense (Q4)

~40 tok/s

~45 tok/s

~50 tok/s

The MoE models are 3-4x faster than the dense ones on your hardware. That's the memory-bandwidth story — decode is bandwidth-bound, and MoEs only activate a fraction of parameters per token.

Tool-Calling Reliability (what matters for OpenCode/Pi)

Model

Tool-Calling Verdict

Source

Qwen 3.6-27B Dense 🏆

Best. Clean tool calls, no wasted thinking tokens. "Handled orchestration cleanly" in real-world video pipeline test at 174K context.

OpenClaw Radar, r/LocalLLaMA

Qwen 3.6-35B-A3B MoE

Good. SWE-bench Verified 73.4%. But can "get lost as tasks require more steps" — MoE routing causes cross-turn coherence to soften on long agent runs.

InsiderLLM, ZoliBen benchmarks

Gemma 4 26B-A4B MoE

Quirky. Forces a reasoning trace by default (thinking field, not content). Community reports "repeatedly entered tool-call loops, wasting tokens on unnecessary reasoning." Fixable with --jinja --chat-template-kwargs '{"enable_thinking":false}' but annoying.

OpenClaw Radar, InsiderLLM

Gemma 4 31B Dense

Good but same thinking-mode quirk. Needs the same flag workaround. Higher quality than 26B MoE on hard reasoning.

Multiple benchmarks

Quality Benchmarks

Benchmark

Qwen 3.6-27B Dense

Qwen 3.6-35B-A3B MoE

Gemma 4 26B-A4B MoE

SWE-bench Verified

77.2%

73.4%

~52%

SkillsBench (agent tasks)

48.2%

28.7%

N/A

Terminal-Bench 2.0

59.3%

51.5%

N/A

LiveCodeBench v6

83.9%

80.4%

N/A

---

My Recommendation for You (32GB VRAM)

You have an embarrassment of riches with 32GB — you can fit any of these comfortably at Q4 with room for context.

🥇 Primary: Qwen 3.6-35B-A3B MoE — The Speed King + Reliable Enough

- ~100-160 tok/s — it feels instant in an agent loop
- Fits in ~22GB at Q4_K_M, leaving 10GB+ for KV cache and context
- Tool calling is good for most tasks (73.4% SWE-bench)
- No thinking-mode quirks, no flag gymnastics
- Apache 2.0 license
-
Caveat: For very long multi-step agent runs (10+ tool calls deep), the MoE routing can drift. For 90% of coding agent work, this is fine.

ollama pull qwen3.6:35b-a3b


🥈 Secondary: Qwen 3.6-27B Dense — For Complex Multi-Step Agentic Tasks

- Slower (~40 tok/s) but most reliable tool calling in the class
-
Best agentic coding scores (SkillsBench +19.5 over the MoE!)
- Rock-solid cross-turn coherence — every token uses every parameter
- Fits in ~17GB at Q4, super comfortable in 32GB
- What Simon Willison called "outstanding for a 16.8GB local model"

ollama pull qwen3.6:27b


🥉 Gemma 4 26B-A4B MoE — If Speed Trumps Everything

- 128-150+ tok/s — the absolute fastest
- The thinking-mode quirk is annoying but solvable
- Tool-call loops are the real issue — some users report it gets stuck in reasoning spirals
- With 32GB VRAM you can also run the 31B dense if you want Gemma quality

The Power Move (With 32GB VRAM You Can)

Run both at once (or swap easily):
-
Gemma 4 26B-A4B MoE for fast chat/quick edits → ~17GB
-
Qwen 3.6-27B Dense for complex multi-step agentic work → ~17GB

Or just pick one based on your dominant workflow:

Your Priority

Pick

Fastest t/s + decent tool calling

Qwen 3.6-35B-A3B MoE

Most reliable tool calling for agents

Qwen 3.6-27B Dense

Maximum speed, willing to work around quirks

Gemma 4 26B-A4B MoE

"I want both speed AND quality"

Run Qwen 3.6-35B-A3B for speed, swap to Qwen 3.6-27B for hard tasks

One More Thing — llama.cpp Settings Matter

For Qwen 3.6 coding: temperature 0.6, top_p 0.95, top_k 20, presence_penalty 0.0

For Gemma 4: ALWAYS include --jinja --chat-template-kwargs '{"enable_thinking":false}' unless you want the reasoning trace.

---

Bottom line for OpenCode/Pi with 32GB VRAM:
I'd go
Qwen 3.6-35B-A3B MoE as your daily driver (fast enough to feel like a cloud API, reliable tool calling for 90%+ of tasks), and keep Qwen 3.6-27B Dense in your back pocket for the complex multi-file refactors where you need every bit of coherence. Both are Apache 2.0, both work with any OpenAI-compatible endpoint, and the community has thoroughly battle-tested them with OpenCode, Aider, and Claude Code proxies.