Disclosure: I maintain Lynkr, the open-source router used in the config examples. The benchmark figures below are third-party or vendor-reported (flagged where vendor-only) — I haven't independently benchmarked these models yet; the point of this post is to help you match models to request classes and test on your own workload.
June 2026 was the busiest month for open-weight coding models in recent memory: GLM-5.2, MiniMax M3, Kimi K2.7 Code, Gemma 4, and NVIDIA's Nemotron 3 Ultra all landed within weeks. If you route your coding agent's simple requests to a local model — the "cloud architect, local coder" pattern — your options just changed meaningfully.
Here's how I'd map the current field onto routing tiers, by hardware budget and by what each model can safely own.
First, the trap: "best open model" ≠ "your local tier"
The headline model of the month, GLM-5.2, scores 62.1% on SWE-bench Pro — above GPT-5.5. It is also a 744B-parameter MoE whose 2-bit quant alone wants ~245 GB of memory. That's an open-weight model, not a local model; for self-hosters it's a $40k-rig proposition (one published build runs it on four RTX PRO 6000s). The same goes for DeepSeek-V4 Pro and MiniMax M3: superb models you'll realistically consume via API, where they belong in your COMPLEX/REASONING tiers, not your local one.
Your local tier is decided by a harsher question: what fits in your VRAM and still makes reliable tool calls?
The local field, by hardware budget
~16 GB RAM (ordinary laptop): Gemma 4 12B. Released June 3 as a dense 12B that genuinely fits consumer RAM (SitePoint's guide). Apache-2.0-class licensing with no usage clauses. This is a SIMPLE-tier model: explanations, one-liners, commit messages, "what does this error mean." I would not hand it an Edit tool.
24 GB GPU (RTX 3090/4090 class): Qwen3.6-27B — still the default answer. The community's consensus "local Claude" since April: within a few points of frontier models on SWE-bench Verified (77.2 reported vs Claude's 80.9 — analysis), Apache-2.0, runs quantized on a single 24 GB card or a ~$2k build. Its known weakness is exactly the one that matters for agents: tool-call reliability drifts in long contexts — fine as a supervised MEDIUM tier, risky as an unsupervised COMPLEX one.
Agentic multi-file edits on similar hardware: Devstral Small 2. Purpose-built for multi-file, tool-driven coding rather than chat (KDnuggets roundup). If your traffic is edit-heavy, it can arguably take MEDIUM-tier mutation requests that I'd keep away from general chat models.
Autocomplete-shaped work: Codestral 22B is fast and good at it — but mind the non-commercial license before using it for work.
One rule that keeps proving out (Pinggy's guide): within the same memory budget, a bigger model at Q4 usually beats a smaller one at Q8. Quantization choice matters nearly as much as model family.
Mapping to tiers
Putting that together into a routing config (Lynkr shown; the mapping logic applies to any router):
# 24 GB GPU + API keys for the hard stuff
TIER_SIMPLE=ollama:gemma4:12b # trivia, explanations, greetings
TIER_MEDIUM=ollama:qwen3.6:27b # code questions, supervised edits
TIER_COMPLEX=deepseek:deepseek-v4-flash # tool-heavy mutations, via API
TIER_REASONING=deepseek:deepseek-v4 # architecture, multi-step planning
Why V4 Flash for COMPLEX: it's the first open-weight model teams report dropping into real agentic pipelines as a frontier substitute on price (OpenRouter's June analysis) — the cheapest "won't break the session" option right now. Kimi K2.7 Code (vendor-reported 58.6% SWE-bench Pro at ~30% fewer reasoning tokens) and GLM-5.2 are strong API-tier alternatives; all the June day-one numbers are vendor-reported, so treat them as directional until LiveBench catches up.
The key discipline: the boundary between MEDIUM and COMPLEX should not be "how big is the request" but "will tools mutate state." Local models in this class handle read-and-explain reliably; exact-match edits and bash execution are where they still break sessions — I wrote up those failure modes here.
What changed vs three months ago
- The floor rose. A 16 GB laptop now runs a genuinely useful SIMPLE tier (Gemma 4). Six months ago that tier meant 3B models that couldn't be trusted with a paragraph.
- The open-weight ceiling now beats proprietary on some coding benchmarks (GLM-5.2 > GPT-5.5 on SWE-bench Pro) — but at server scale, which strengthens the hybrid pattern: open models via cheap APIs up top, small open models on your metal below.
- MoE won. Every serious June release is Mixture-of-Experts. For self-hosters this cuts both ways: better quality-per-active-param, but total memory footprints that keep the top tier out of reach.
- Licensing is consolidating around MIT (DeepSeek) and Apache-2.0 (Qwen, Gemma) for the models you'd actually build on.
Test on your traffic, not on benchmarks
Every number above is someone else's workload. The honest way to pick your local tier: route a week of your real traffic through whatever candidates fit your hardware, and count session survival — how often the local model's tool calls held up — not just benchmark deltas. That's a one-line config change per candidate, and your own telemetry will contradict at least one thing this post told you.
Lynkr is Apache-2.0, self-hosted, and treats every model above as a first-class routing tier: github.com/Fast-Editor/Lynkr.
Top comments (0)