Qwen 3.8 vs Ornith 1.5: A Real-World Comparison of Two 35B-Class Local Models

Two Big Drops in Five Days: A Head-to-Head in the Local LLM Space

The local LLM scene saw two major releases in the span of five days. On August 14th, Alibaba dropped the dense model Qwen 3.8 (27B); just five days later, on August 19th, Deep Green Force followed with the open-source Ornith 1.5 (35B). Both are open-weight, both target consumer GPU users — in other words, both are coming straight for the card sitting in your PC.

Which one actually wins is far from clear-cut. In a direct coding benchmark, Qwen comes out ahead decisively: 24 out of 30 tasks passed versus Ornith's 15 — a seemingly lopsided result. But when both models were asked to build a complete 3D game from scratch, the script flipped immediately — Qwen delivered with a single prompt and zero hand-holding, while Ornith needed six rounds of pushing before it could barely catch up. And once the context window gets long enough, the speed advantage reverses yet again.

This article is adapted from a 35-minute full comparison video on the YouTube channel PandaMakingMoney. There are no sponsors, no financial ties to either model's backers, and all conclusions are based on public benchmarks, independent testing, and feedback from users who actually ran these models on their own hardware. Bottom line up front: who wins depends entirely on what you're optimizing for.

Architecture Comparison: MoE Saves Compute vs. Dense Saves VRAM

Ornith 1.5: The Compute Edge (and VRAM Cost) of Mixture-of-Experts

Ornith 1.5 (35B) comes from Deep Green Force under an MIT license — free to modify and use commercially. The family also includes a lightweight 9B variant for low-end hardware and a 397B flagship (completely out of reach for home machines), making the 35B the real workhorse for consumer GPUs.

It's a Mixture-of-Experts (MoE) model: internally it has 256 experts, and for each incoming token, the router selects only 8 to activate while the other 248 sit idle. Total parameters are 35B, but only roughly 3B are activated per token — which is exactly why it runs fast at inference.

The catch, however, is real: routing is a dynamic per-token decision, and you can't predict which experts will be needed next, so all 35B parameters must reside in VRAM simultaneously. Compute is cheap; VRAM is not.

Qwen 3.8: Dense Architecture with Hybrid Attention

Qwen 3.8 (27B) takes a different approach — a dense model under the Apache 2.0 license from Alibaba's Qwen team. All 27.7B parameters are activated every time, with no routing, no expert selection, and the entire model working on every token.

But it applies its efficiency to the attention mechanism instead — hybrid attention. Of the model's 64 layers, only 16 use traditional full attention (which requires tracking an ever-growing conversation history), while the remaining 48 use linear attention with a fixed state size (DeltaNet). The result is that KV cache growth is nearly flat as context expands: only about 0.5GB more at 8K context, and only about 2GB more at 32K — a significant advantage for long coding sessions and extended conversations. VRAM scales well, but every token activates the full model.

Both companies follow the same playbook: flagship models live in the cloud (Ornith's 397B, Qwen's 2.4-trillion-parameter Max via API), while the "little siblings" are what you actually get to self-host. Two very different engineering bets, now openly on the table.

Benchmark Verification: Vendor-Reported Numbers vs. Independent Evaluations

Most comparison videos fall apart by just pasting vendor-reported numbers into a table and calling it a day. Here, we apply a healthy discount.

On Ornith's side, Deep Green Force self-reports Terminal Bench 2.1 scores in the low 60s and their own coding leaderboard in the low 70s. That's impressive for the 35B class — but every number comes from the vendor testing their own model, in their own evaluation environment, on their own hardware, with no third-party involvement. That doesn't mean the numbers are fabricated, but they should be read with the same skepticism you'd give a car manufacturer claiming their car is the fastest.

Qwen 3.8's self-reported gains are even more dramatic: Terminal Bench 2.1 jumps from 63.4 to 73.0, Deep SWA nearly triples from 13.3 to 42.2, and OS World validation score leaps from 63.9 to 84.3. Also self-reported.

The only independent evidence in this entire comparison comes from Artificial Analysis — a third party with no financial ties to Alibaba. They gave Qwen an intelligence index of 52, matching full-reasoning GPT-5.6, with an agentic score of 51 that edges out Claude Opus 4.8, which launched less than three months ago. The tradeoff is verbosity: Qwen generated roughly 160 million tokens in the same evaluation, compared to a median of 43 million for open models of similar size — nearly 4× more. Ornith, meanwhile, has yet to receive any independent verification.

Real-World Testing: Coding Ability and Autonomy Under Stress

Round 1: 30 One-Shot Coding Tasks

A community member designed 30 one-shot coding tasks using Three.js — one attempt per task, no retries, no additional guidance. This removes the safety net of iteration and more closely mirrors real automated coding workflows. Both models were run on Q6 quantization to keep conditions fair.

Qwen 3.8 passed 24 out of 30 — the highest in the field. Of Ornith's 15 passes, several were blank screens or outputs that technically met the scoring criteria but weren't actually functional — they counted as passing, but wouldn't work in practice. To be fair, though: on the tasks where Ornith did succeed, the output quality was noticeably better than its predecessor. Overall, Ornith burned 20–40% more tokens than Qwen, some of which were spent iterating toward final outputs that still didn't land — paying compute costs for results that never materialized.

Round 2: Building a GTA-Style 3D Game From Scratch

The second test was a pure real-world workflow. The tester ran a coding agent on an Apple M3 Ultra with 512GB of unified memory — enough that VRAM was never a bottleneck — to isolate the model's own capabilities. The task was ambitious: a single natural-language prompt asking for a playable 3D game inspired by GTA: San Andreas, with a large open world, the ability to steal and drive vehicles, decent visuals, and everything generated from scratch.

Qwen 3.8 completed the task in one attempt with zero additional guidance — a strong signal for autonomous coding agent use cases. Ornith required six rounds of iteration, with the tester describing it as "constantly pushing and pulling," concluding that it "is simply not at the same level of autonomy as Qwen."

Long Context Performance: The Speed Reversal

But in the same test, the speed story flipped completely. As the conversation extended into very long contexts, Ornith's throughput gradually declined from 75 tokens/sec to around 45 — but remained usable well past the 160K token mark. Qwen, upon crossing the same threshold, dropped to roughly 5 tokens/sec — effectively stalled. This isn't a minor gap; at long context lengths, one model is still usable and the other is functionally broken.

VRAM Requirements and Buying Recommendations

How Each Model Fits Different VRAM Tiers

Ornith's recommended quantization is Q4_K_M, which requires around 23GB for the model plus runtime overhead to run comfortably — official and community guidance suggests budgeting 27GB or more. The MoE architecture becomes a liability here: all 256 experts must live in VRAM simultaneously, even if only 8 are used per token. The fallback for smaller cards is CPU offloading some experts to system RAM — a 12GB card can technically run it, but the speed penalty is immediately visible.

Qwen at the same Q4_K_M quantization only needs 16–19GB, and a 24GB card leaves room for 56K–64K tokens of context — thanks to those 48 linear attention layers. Breaking it down by GPU tier:

• 24GB VRAM: Qwen is comfortable; Ornith is running at the edge
• 16GB VRAM: Qwen is clearly the better fit
• 12GB VRAM: Ornith requires CPU offloading or aggressive quantization
• 32GB+: Both models can run freely at Q6/Q8

Final Recommendations

Choose Qwen 3.8 if: you're working in coding agent workflows and need the model to operate with minimal hand-holding — its single-pass performance is far ahead, it's the only option with independent third-party validation (Artificial Analysis score of 52), it fits more comfortably on 16–24GB cards, and its toolchain (including Unsloth dynamic quantization) is the most mature.

Choose Ornith 1.5 if: your use case requires an MIT license rather than Apache 2.0 for commercial reasons, you're running extended long-context conversations over time, and you have the VRAM budget to accommodate it.

A Note on Independent Verification

Ornith 1.5 has only been out for a week, and nearly all its benchmark scores are vendor-disclosed. Independent third-party validation at the 35B tier is still very thin, and there have even been skeptics on Hacker News questioning whether "self-improvement" at this scale is a credible claim. This isn't a verdict on its quality — the engineering is solid and the speed advantage is real — but trust is earned through independent verification over time, not through a launch-week news cycle. Right now, Qwen has traveled further down that road. This is a gap in evidence, not a judgment on capability.