Local LLM talk is busy on X. We counted about 9,300 posts on local models and Ollama in our sample. Two open coding models arrived within a week: Ornith-1.5-122B (on Hugging Face from Oct 16) and Recursive's Flint (Oct 21). Both post big coding numbers. Here's how they compare, and what you'd need to run them.
At a glance
| Flint | Ornith-1.5-122B | |
|---|---|---|
| Developer | Recursive (RecursiveSI) | Ornith (ornith-ai) |
| Architecture | MoE, 39.8B total, ~2.5B active | MoE, 122B |
| Context | 262,144 tokens | 262,144 tokens (≈1M with YaRN ×4) |
| License | Apache 2.0 | MIT |
| SWE-bench Verified | 74.2% | 84.5 (non-hack) |
| SWE-bench Pro | 89.7% (Pro v2) | 76.9 (no-hack) |
| Terminal-Bench 2.1 | 54.9% | 81.2 |
| Smallest official quant | ~23 GB (NVFP4 GGUF) | ~73 GB (Ollama) |
A note on the table: Flint's "SWE-bench Pro v2" and Ornith's "SWE-bench Pro (no-hack)" are labelled differently, so don't compare those two numbers directly. Every score here comes from the developers' own evaluations.
Flint: small active size, agent-focused
Recursive describes Flint as a mixture-of-experts model "for coding agents", built to investigate codebases, fix bugs, implement features and write tests. It has 39.8B total parameters, but only about 2.5B are active per token, so each step is cheap to compute.
What you get:
- A BF16 checkpoint of about 80 GB, plus an SGLang serving kit. Recursive says it runs on one B300 or H200 GPU per replica. Transformers support isn't included yet.
- Quantized GGUFs:
flint-int4-gguf(≈26 GB) andflint-nvfp4-gguf(≈23 GB). Recursive says quantized variants are prepared separately for NVIDIA and Apple silicon runtimes. - Recommended settings: temperature 1.0, top-p 0.95, top-k 20, and an output budget of at least 32,768 tokens for hard problems.
Recursive's own caveat is worth repeating: results "can vary with checkpoint precision and inference runtime". A 4-bit build won't necessarily match the BF16 scores.
Ornith 1.5: bigger, higher scores, more memory
Ornith-1.5-122B is the largest member of a family that also includes 9B, 35B and 397B models on Ollama. Its card reports 84.5 on SWE-bench Verified (non-hack), 76.9 on SWE-bench Pro (no-hack), 81.2 on TerminalBench 2.1 and 80.0 on CyberGym.
The card says the 122B "serves on a single 80GB GPU". Read that with the file sizes in mind:
- The BF16 GGUF is about 249 GB.
- Q4_K_M is about 76 GB on Hugging Face. Ollama's 122B tags are 73–77 GB.
So "single 80 GB GPU" in practice means a 4-bit-class build with little memory left over for long contexts. Serving needs vLLM ≥ 0.19.1 or SGLang ≥ 0.5.9. Context is 262,144 tokens natively, and the card says YaRN scaling (factor 4.0) stretches it to roughly 1M.
The quickest way to try it:
ollama run ornith-1.5:122b
If that's too big, ornith-1.5:35b (23 GB) and ornith-1.5:9b (6.6 GB) are on Ollama too. Our VRAM guide explains how model size, quantization and context length add up.
Which fits your machine?
| Your hardware | Realistic option |
|---|---|
| 8–16 GB GPU or Mac | Ornith 1.5 9B, or a cloud API |
| 24–32 GB GPU, 32–48 GB Mac | Flint INT4/NVFP4 GGUF (~23–26 GB), Ornith 35B |
| 80 GB GPU, 96–128 GB Mac | Ornith 122B 4-bit builds |
| B300/H200 | Flint BF16 via Recursive's SGLang kit |
These are rough guides based on file sizes. You need headroom beyond the file for context (KV cache) and your OS. Our local AI hardware sizing sheet does the maths, and if you'd rather rent, see what it costs to rent an H100 or B300.
Plugging them into a coding agent
Both serve an OpenAI-compatible API (vLLM, SGLang, llama.cpp or Ollama), so any agent that accepts a custom base URL can use them. Our step-by-step OpenCode guide for Ollama, LM Studio and llama.cpp shows the config, and Ollama for beginners covers the basics.
Bottom line
- Want maximum benchmark score and have 80 GB+? Try Ornith 122B.
- Want something that runs on one high-end consumer card? Flint's quantized builds are the more realistic choice.
- Either way: run a week of your own tasks. Vendor scores often don't match daily use.
